XLRI Jamshedpur | Data Analytics for Finance | Raman

40-Hour Module Map

Mod Title Hours Primary Career Application
ule
1 Foundations — The Financial Analyst's Data 5 All roles
Toolkit
2 Financial Statement & Ratio Analytics at Scale 5 Equity Research, Credit, FP&A
3 Valuation Analytics — Building Data-Driven 6 IB, PE, Equity Research
Models
4 Market & Equity Analytics 5 Equity Research, Asset Management
5 Credit & Risk Analytics 5 Banking, Credit Risk, NBFCs
6 M&A and Deal Analytics 5 IB, PE, Corporate Development
7 Predictive & Machine Learning Applications in 6 Quant, Risk, FinTech, PE Screening
Finance
8 Capstone — The Analyst's Deliverable 3 Integration + Client-Ready Output

How Every Chapter Is Built

Element Purpose
Concept, Plainly Stated What it is and why a finance professional needs it — before any formula
The Mechanics The formula, SQL query, or Python code required to execute it
Worked Example A fully solved problem using real or realistic Indian financial data
Case Lab A mini-assignment mirroring a genuine deal, research, or credit task
From Raman's Desk A war story or judgment note — the thing that separates seniors from
juniors
Analyst's Caution Where this technique commonly goes wrong in practice

For Academic Use Only | Page 1

XLRI Jamshedpur | Data Analytics for Finance | Raman

MODULE 1: FOUNDATIONS — THE FINANCIAL ANALYST'S DATA

TOOLKIT

1.1 Why Data Analytics Failed Finance Before — And Why It Won't Now
Before we open a spreadsheet, you need to understand a piece of history I lived through. In 2007 and 2008,
some of the most sophisticated quantitative models ever built inside investment banks — models built by
people with PhDs in mathematics — told their risk committees that mortgage-backed securities were safe. The
models were not wrong on their own terms. They were fed historical data that never included a nationwide
housing price decline, so they concluded a nationwide housing price decline was nearly impossible. This is
Garbage In, Garbage Out (GIGO), and it is the single most important warning in this entire course.

Every technique you learn in this book is powerful. Every technique in this book can also be catastrophically
wrong if you do not understand the data feeding it, the assumptions embedded in it, and the limits of what it
can tell you. I am not teaching you analytics to make you trust models more. I am teaching you analytics to
make you interrogate them better.

FROM RAMAN'S DESK

At JPMorgan, I once had a first-year analyst hand me a beautifully formatted DCF model with a 45%
IRR on a mid-market acquisition. It looked perfect. It took me four minutes to find that he had pulled a
working capital figure from the wrong fiscal year in the data feed — a simple date-range error in an
Excel pull. The model was elegant. The underlying number was wrong. Nobody in the room but me
checked the raw data before trusting the output. From that day forward, my rule for every analyst
who worked for me was simple: before you trust your model, trace every number in it back to its
source, by hand, at least once.

1.2 Excel as an Analytical Engine, Not a Data Entry Tool

Most MBA students use perhaps 5% of Excel's actual analytical capability. In this course, Excel is not a place to
type numbers — it is the fastest tool you will ever have for a first-pass analysis of a mid-sized dataset, and in
many boutique advisory firms and family offices, it remains the primary tool used for live client work.

1.2.1 Advanced Formulas Every Financial Analyst Must Master

Function Financial Use Case
INDEX-MATCH / XLOOKUP Pulling specific line items from multi-year, multi-company financial
statement pulls
SUMIFS / COUNTIFS / AVERAGEIFS Building peer-set averages, conditional ratio calculations across a
screened universe
OFFSET / INDIRECT Building dynamic, rolling time-series ranges for trend charts that update
automatically
Data Tables (What-If Analysis) One- and two-variable sensitivity tables for DCF assumptions (WACC vs
terminal growth)
Array Formulas / SUMPRODUCT Multi-condition weighted calculations — e.g., portfolio-weighted return
across positions

For Academic Use Only | Page 2

XLRI Jamshedpur | Data Analytics for Finance | Raman

TEXT functions Cleaning messy company names, ISINs, and identifiers pulled from
(LEFT/RIGHT/MID/TRIM) different data vendors

1.2.2 Power Query — The Most Underused Tool in Finance

Power Query is Excel's built-in data transformation engine, and in my experience fewer than one in ten MBA
graduates has ever opened it. It allows you to connect to a data source — a folder of quarterly filings, a CSV
export from Bloomberg, a database table — clean it, reshape it, and refresh it with one click when new data
arrives. In practice, this is how you build a repeatable pipeline instead of redoing the same manual cleanup
every quarter.

WORKED EXAMPLE — Building a Repeatable Peer-Set Pipeline

Imagine you cover 25 companies in the Indian cement sector for equity research. Every quarter, you
need updated financials for all 25. Without Power Query: you manually download 25 files, copy-paste
into a master sheet, and fix formatting errors — roughly 3-4 hours of work, every quarter, prone to
copy-paste errors. With Power Query: you build the transformation logic once — connect to the
folder, unpivot the data, standardize the line items, append into one table. Each subsequent quarter,
you refresh in under two minutes. This is the difference between an analyst who works late every
quarter and one who has time left over to actually think about what changed.

1.3 SQL for the Finance Professional

You do not need to be a database administrator. You need to be able to ask a database a precise question and
get back exactly the rows you need — nothing more. In my career, this single skill separated analysts who
could self-serve their own data from analysts who waited days for someone else to pull it for them.

1.3.1 The Four Queries That Cover 90% of Finance Use Cases
Query Pattern Finance Application Core Syntax
SELECT with WHERE filtering Pulling all companies in a sector above SELECT * FROM companies WHERE
a market cap threshold sector='Cement' AND market_cap >
5000
JOIN across tables Combining a company master table SELECT [Link], [Link] FROM
with a financials table by company ID companies a JOIN financials b ON
[Link]=b.company_id
GROUP BY with aggregation Computing sector-average margins SELECT sector, AVG(ebitda_margin)
across a screened universe FROM financials GROUP BY sector
Window functions (RANK, Ranking companies by ROE within SELECT name, revenue, LAG(revenue)
LAG) their sector; computing YoY growth OVER (PARTITION BY company_id
ORDER BY year) FROM financials

WORKED EXAMPLE — SQL for Credit Portfolio Screening

A bank's credit risk team maintains a loan book database with a `loans` table (loan_id, borrower_id,
sector, outstanding_amount, days_past_due) and a `borrowers` table (borrower_id, name,
credit_rating). To find every loan over Rs. 10 crore that is more than 90 days past due, in a sector the
bank has flagged as high-risk: SELECT [Link], l.outstanding_amount, l.days_past_due FROM loans l
JOIN borrowers b ON l.borrower_id = b.borrower_id WHERE l.outstanding_amount > 100000000 AND
l.days_past_due > 90 AND [Link] IN ('Real Estate','Infrastructure') ORDER BY l.outstanding_amount

For Academic Use Only | Page 3

XLRI Jamshedpur | Data Analytics for Finance | Raman

DESC; — This single query, which takes seconds to write once you know the pattern, replaces what
used to take a junior analyst an entire day of manual Excel filtering across multiple exported reports.

1.4 Python for Finance — The Minimum Viable Toolkit

I am not going to teach you to build software. I am going to teach you the narrow slice of Python that a
working finance professional actually uses: pandas for tabular data manipulation, NumPy for numerical
computation, and just enough syntax to read, clean, and analyze a financial dataset larger or messier than
Excel can comfortably handle.

1.4.1 Core pandas Operations for Financial Data

Operation Code Pattern Financial Use
Reading data df = pd.read_csv('[Link]') Loading a multi-company financial
statement export
Filtering rows df[df['revenue'] > 1000] Screening a universe by a financial
threshold
Grouping and aggregating [Link]('sector')['roe'].mean() Sector-average ratio computation across
hundreds of companies
Merging datasets [Link](df1, df2, Joining financials with market price data
on='company_id') by company identifier
Calculating rolling metrics df['revenue'].pct_change() Computing period-over-period growth
rates across a time series
Pivoting df.pivot_table(index='company', Reshaping long-format financial data
columns='year', values='ebitda') into a year-over-year comparison table

FROM RAMAN'S DESK

A managing director does not care whether your code is elegant. A managing director cares whether
the number is right and whether you can explain, in one sentence, how you got it. I have seen brilliant
coders fail in finance because they could produce a beautiful analysis nobody in the room could
interpret in the two minutes available before the next agenda item. Write code that a colleague could
read in thirty seconds and understand what it does. That discipline will serve you longer than any
advanced technique.

1.5 Data Sourcing in Indian and Global Markets

Knowing where to find authoritative, defensible data — and knowing which sources to trust for which purpose
— is itself a professional skill. I have seen deals delayed because an analyst used unreliable secondary data
instead of the primary regulatory filing.

Source Best Used For Notes

Bloomberg Terminal Real-time market data, bond pricing, Industry standard at IBs and asset
cross-asset analytics, consensus managers; expensive but authoritative
estimates
CMIE Prowess Deep historical Indian company Widely used in Indian equity research
financials, ownership data and academic finance research
Capitaline Indian company financials, corporate Common in Indian brokerage and

For Academic Use Only | Page 4

XLRI Jamshedpur | Data Analytics for Finance | Raman

actions, industry data merchant banking research teams

NSE/BSE Bhavcopy Daily raw price, volume, and delivery Free, official, and the base data feed
data for listed Indian securities for most Indian quant strategies
[Link] Quick-access Indian company financials Excellent for coursework and quick
and ratios for screening screening; always verify against filings
for real work
Company Annual Reports / The single source of truth for any Always the final verification step — I
BSE-NSE Filings number that matters in a live deal have never closed a deal without
checking primary filings
RBI / SEBI Publications Macro data, sector-level regulatory Essential for credit risk and macro-
data, banking system statistics driven strategy work

ANALYST'S CAUTION: A number from a data vendor is a starting point, never a conclusion. Every material
figure that goes into a client deliverable — a valuation, a credit recommendation, an investment memo —
must be traceable to a primary source: the company's own filing, the regulator's own publication, the
exchange's own data feed. I have seen careers damaged by a single unverified number in a client
presentation.

Module 1 — Self-Assessment
Q# Question
1 Explain, using the 2008 financial crisis as your example, what Garbage In Garbage Out means for a finance
professional building a quantitative model.
2 Design a Power Query pipeline (in words — steps, not code) that would let you refresh a 30-company peer
set's financials every quarter with one click.
3 Write a SQL query that returns the top 10 companies by revenue growth within a specific sector, using a
companies table and a financials table.
4 Using pandas syntax, describe how you would compute the average ROE by sector across a 500-company
dataset.
5 You are asked to value a private company for an acquisition. List, in priority order, the data sources you
would use and explain why primary filings outrank all vendor data.

For Academic Use Only | Page 5

XLRI Jamshedpur | Data Analytics for Finance | Raman

MODULE 2: FINANCIAL STATEMENT & RATIO ANALYTICS AT SCALE

2.1 Automating Financial Statement Extraction and Standardization

The single biggest time cost in equity research and credit analysis is not the analysis — it is getting clean,
comparable data in the first place. Every company reports slightly differently. One classifies a cost as 'other
operating expense,' another buries the same cost inside 'cost of goods sold.' Before you can compare 30
companies, you must standardize their statements onto a common template.

2.1.1 The Standardization Framework

Step What Happens Why It Matters
1. Map to a common Every company's line items are Without this, 'Revenue' from Company A
chart of accounts reclassified into a fixed, standard set of and 'Net Sales' from Company B may not be
categories comparable
2. Adjust for one-off Exceptional gains/losses, impairments, A one-time asset sale can distort a margin
items and restructuring costs are separated comparison if left in reported EBITDA
from core operations
3. Align fiscal periods Companies with different year-ends are Comparing a March year-end company to a
aligned to comparable trailing-twelve- December year-end company directly can
month windows mislead
4. Currency and unit All figures converted to a common Essential when peer sets include ADR-listed
normalization currency and unit (e.g., Rs. crore) or multinational comparables

FROM RAMAN'S DESK

At Morgan Stanley, we maintained standing templates for every sector we covered — cement,
banking, IT services, pharma — because each sector has its own quirks in how companies report. A
bank's 'revenue' is fundamentally different from a manufacturer's. My first instruction to any new
analyst joining a sector coverage team was always: spend your first two weeks not building models,
but building the standardization template. Every hour invested there saves ten hours later.

2.2 Ratio Analytics Across Large Peer Universes

A single company's ratios tell you a little. The same ratios computed across an entire peer universe, ranked
and benchmarked, tell you a great deal — this is where analytics genuinely outperforms manual analysis,
because no human can hold fifty companies' relative positioning in their head.

2.2.1 The Four Ratio Families

Category Key Ratios Formula What It Reveals
Liquidity Current Ratio, Quick Ratio CA/CL ; (CA-Inventory)/CL Can the company meet
near-term obligations?
Leverage Debt-Equity, Interest Coverage Total Debt/Equity ; How much financial risk
EBIT/Interest has the company taken
on?
Profitability Gross Margin, EBITDA Margin, Various — margin = How efficiently does the

For Academic Use Only | Page 6

XLRI Jamshedpur | Data Analytics for Finance | Raman

ROE, ROCE profit/revenue; ROE = company convert

PAT/Equity revenue and capital into
profit?
Efficiency Asset Turnover, Inventory Revenue/Assets ; How well does the
Days, Receivable Days 365×(Inventory/COGS) ; company use its working
365×(Debtors/Revenue) capital and assets?

WORKED EXAMPLE — Computing Ratios Across a 40-Company Peer Set (pandas)

Suppose you have standardized financials for 40 Indian IT services companies in a DataFrame `df` with
columns: company, revenue, ebitda, net_profit, total_debt, total_equity, interest_expense. To
compute EBITDA margin and ROE for every company in one line each: df['ebitda_margin'] =
df['ebitda'] / df['revenue']; df['roe'] = df['net_profit'] / df['total_equity']. To then rank all 40
companies by ROE within the sector: df['roe_rank'] = df['roe'].rank(ascending=False). In under five
lines of code, you have what would take an analyst hours to compute manually across 40 separate
company files — and, critically, you have it in a form you can instantly re-rank, filter, or visualize.

2.3 Trend and Variance Analytics

A single quarter's ratio is a snapshot. The trend across eight to twelve quarters is where the real signal lives. I
have caught more early warning signs of financial distress from a trend line than from any single period's
numbers.

2.3.1 What a Trend Break Tells You

Trend Signal What It May Indicate What to Check Next
Receivables days rising for 3+ Weakening customer collections; Cross-check against cash flow from
consecutive quarters possible revenue recognition operations trend — is cash following
aggression; channel stuffing reported profit?
Gross margin declining while Pricing pressure, rising input costs, or Segment-level disclosure,
revenue grows a mix shift toward lower-margin management commentary, input
products cost indices
Inventory days rising sharply Demand slowdown, obsolescence Compare against sector peers — is
risk, or overproduction this company-specific or sector-
wide?
Interest coverage ratio Rising financial distress risk; covenant Debt maturity schedule, refinancing
declining toward 1.5x or below breach risk approaching risk, covenant terms in debt filings
Operating cash flow diverging Possible earnings quality issue — Working capital movement analysis;
negatively from net income profit not converting to cash look for aggressive revenue
recognition

FROM RAMAN'S DESK

I once flagged a mid-cap Indian company to a credit committee eighteen months before its public
default, purely from a trend chart. Nothing in any single quarter's numbers looked alarming — every
individual ratio was within a defensible range. But receivables days had crept up for six straight
quarters while operating cash flow diverged further and further below reported net income each
period. No single number was the smoking gun. The trend was. This is the entire argument for why
trend analytics matters more than point-in-time ratios in credit and equity work.

For Academic Use Only | Page 7

XLRI Jamshedpur | Data Analytics for Finance | Raman

2.4 Peer Benchmarking and Percentile Analysis

Raw ratios are meaningless without context. A 15% EBITDA margin is excellent in low-margin retail and
mediocre in software. Percentile ranking within a defined peer set converts an absolute number into a
relative, decision-useful signal.

WORKED EXAMPLE — Building a Sector Scorecard

For a 25-company Indian cement sector peer set, you compute five key metrics for each company:
EBITDA margin, ROCE, Debt/EBITDA, Revenue growth (3-yr CAGR), and Receivable days. For each
metric, you rank every company into percentiles (0-100) versus the peer set. A company scoring in the
85th percentile on ROCE and EBITDA margin, but the 15th percentile on Debt/EBITDA (i.e., very low
leverage, meaning conservative balance sheet) presents a very different investment or credit picture
than a company in the 85th percentile on all metrics including leverage. The scorecard format — a
simple heat-mapped table — is precisely how I would open an Investment Committee discussion: not
with fifty pages of analysis, but with one table that instantly shows where a company sits relative to
its true peer set.

2.5 Case Lab — Detecting Financial Stress Before It's Public

CASE LAB — Detecting Financial Stress Before It's Public
You are given eight quarters of standardized financial data for a mid-sized Indian NBFC, including
revenue, net profit, receivables, operating cash flow, total debt, and interest expense — pulled from
public filings. Using the trend and ratio analytics techniques from this module: (1) Compute the
trailing trend for receivables days, interest coverage, and the operating-cash-flow-to-net-income ratio
across all eight quarters. (2) Identify which quarter, if any, marks an inflection point where the trend
changes character. (3) Build a one-page memo — the kind I would have expected on my desk the
morning after a credit committee request — stating whether you would flag this company for
enhanced monitoring, and precisely which three metrics drove your conclusion. (4) Present your
findings to the class in under three minutes: this is the actual time constraint I operated under in real
credit committee meetings.

ANALYST'S CAUTION: A single bad ratio rarely predicts distress reliably — companies can explain away one
weak quarter convincingly, and often correctly. It is the persistence and co-movement of multiple
deteriorating signals across consecutive periods that separates a genuine early warning from noise. Never flag
a company on one data point alone.

Module 2 — Self-Assessment
Q# Question
1 Explain why standardizing a chart of accounts across companies is necessary before any peer comparison
is valid. Give one concrete example of a line item that is commonly classified differently across companies.
2 A company's EBITDA margin has stayed flat for four quarters, but its operating cash flow has declined by
30% over the same period. What are three possible explanations, and what additional data would you
request to distinguish between them?
3 Design a percentile-based sector scorecard (list the metrics you would include and explain your choice) for
benchmarking 20 Indian pharmaceutical companies.
4 Using pandas syntax, write the code to compute a 4-quarter rolling average of receivables days for each
company in a dataset with columns company, quarter, and receivable_days.

For Academic Use Only | Page 8

XLRI Jamshedpur | Data Analytics for Finance | Raman

5 What is the danger of flagging a company as financially distressed based on a single quarter's ratio
deterioration? How does trend analysis mitigate this risk?

For Academic Use Only | Page 9

XLRI Jamshedpur | Data Analytics for Finance | Raman

MODULE 3: VALUATION ANALYTICS — BUILDING DATA-DRIVEN

MODELS

3.1 From Static Models to Dynamic, Data-Fed Valuation

The DCF model you learned in your core Corporate Finance course is correct in its mechanics. What it likely
lacked was a data architecture — a way to feed live market data, updated financials, and comparable company
multiples into the model without manually retyping numbers every time something changes. In a live deal,
numbers change daily. A model that requires manual re-entry every time is a model that will eventually
contain an error.

Static Model Approach Data-Fed Model Approach

Analyst manually types WACC inputs, comps multiples, Model pulls risk-free rate, beta, and comps multiples
growth rates from a linked data source, refreshed on demand
Comp set is a fixed list, manually updated when Comp set is generated by a screening query against
someone remembers live market data (sector, size, geography filters)
Sensitivity analysis is manually recalculated for each Data Tables or Monte Carlo simulation automatically
scenario generate the full sensitivity surface
Errors from manual re-entry are common and hard to Errors are isolated to the data source or
trace transformation logic, and easier to audit

FROM RAMAN'S DESK

Early in my career at a merchant bank, I inherited a colleague's DCF model mid-deal after he left the
firm. It took me two full days just to figure out which cells were hardcoded assumptions and which
were meant to be linked formulas — because half of what looked like a formula was actually a
manually pasted value from three weeks earlier. That experience is why, to this day, I insist every
model I sign off on clearly separates: raw data inputs, assumptions, and calculated outputs, in three
visually distinct sections. This is not a stylistic preference. It is a professional discipline that prevents
exactly the kind of error that once cost my colleague's successor two wasted days.

3.2 Comparable Company Analysis at Scale

A comp set of five hand-picked companies is common in student projects. A comp set of thirty to fifty
companies, screened, cleaned, and statistically summarized, is what a real banking or PE analyst is expected to
produce — because a small, hand-picked set is far more vulnerable to selection bias, whether intentional or
not.

3.2.1 Building and Cleaning a Comp Set

Step What Happens Common Pitfall
1. Screen the Filter by sector, geography, size band, Casting too wide a net includes companies that
universe and business model similarity are not truly comparable in risk or growth
profile
2. Pull trading EV/EBITDA, EV/Revenue, P/E for each Mixing trailing and forward multiples without
multiples company, computed consistently labeling clearly which is which

For Academic Use Only | Page 10

XLRI Jamshedpur | Data Analytics for Finance | Raman

3. Remove Companies with distressed, negative, or Silently deleting inconvenient outliers without
statistical outliers abnormal multiples are flagged and disclosure is a serious analytical and ethical
typically excluded or footnoted error
4. Compute Mean, median, and quartile range of Reporting only the mean when the distribution
summary statistics multiples across the clean comp set is skewed — median is often more robust

WORKED EXAMPLE — Comp Set Construction in pandas

You have pulled trading data for 60 Indian mid-cap manufacturing companies into a DataFrame `df`
with columns: company, sector, ev, ebitda, market_cap. First, compute EV/EBITDA for all:
df['ev_ebitda'] = df['ev'] / df['ebitda']. Remove negative or extreme outliers: clean = df[(df['ev_ebitda']
> 0) & (df['ev_ebitda'] < df['ev_ebitda'].quantile(0.95))]. Then get summary statistics for your target's
sub-sector: clean[clean['sector']=='Auto Components']['ev_ebitda'].describe() — this returns count,
mean, standard deviation, and quartiles in one line, giving you a defensible statistical basis for your
multiple range rather than a single cherry-picked comparable.

3.3 Sensitivity and Scenario Analytics

A single-point DCF value is almost always wrong — not because the model is flawed, but because it represents
one specific combination of assumptions out of a very wide plausible range. The professional standard is to
present a range, built from systematic sensitivity analysis, not a single number presented with false precision.

3.3.1 Two-Variable Data Tables

Excel's Data Table feature allows you to see how your valuation output (Enterprise Value, Equity Value per
Share) changes across a grid of two varying inputs simultaneously — most commonly WACC on one axis and
terminal growth rate on the other.

WORKED EXAMPLE — WACC vs Terminal Growth Sensitivity Table

A DCF model produces a base case Enterprise Value of Rs. 4,200 crore at WACC = 11% and terminal
growth = 4%. Building a Data Table across WACC (10%, 10.5%, 11%, 11.5%, 12%) and terminal growth
(3%, 3.5%, 4%, 4.5%, 5%) produces a 5x5 grid of 25 possible Enterprise Values, ranging perhaps from
Rs. 3,400 crore (high WACC, low growth) to Rs. 5,600 crore (low WACC, high growth). Presenting this
grid — not just the single base-case number — is what separates a defensible valuation opinion from
an overconfident one. In my experience, Investment Committees trust a well-reasoned range far more
than a suspiciously precise single figure.

3.3.2 Monte Carlo Simulation for DCF Assumptions

Where a two-variable Data Table shows sensitivity to two inputs, Monte Carlo simulation lets you assign
probability distributions to many inputs simultaneously — revenue growth, margin trajectory, WACC, terminal
growth — and run thousands of randomized combinations to produce a full probability distribution of possible
valuation outcomes, not just a grid.

WORKED EXAMPLE — Monte Carlo Simulation Setup (conceptual, Python)

Using NumPy, you define each key assumption as a probability distribution rather than a fixed
number: revenue_growth = [Link](0.08, 0.02, 10000) generates 10,000 random draws
from a normal distribution centered at 8% growth with 2% standard deviation. Similarly for margin

For Academic Use Only | Page 11

XLRI Jamshedpur | Data Analytics for Finance | Raman

and WACC. You then run your DCF formula across all 10,000 combinations, storing each resulting
valuation. The output is a full distribution: you can now state, for example, 'there is a 90% probability
the Enterprise Value falls between Rs. 3,600 crore and Rs. 5,100 crore' — a materially more honest
and useful statement to an Investment Committee than a single-point estimate.

3.4 Precedent Transaction Analytics

Precedent transactions — prices paid in past comparable M&A deals — are a critical valuation cross-check,
particularly for control premiums. Building a queryable database of past deals, rather than relying on memory
or scattered files, is standard practice at every investment bank I worked at.

WORKED EXAMPLE — Querying a Precedent Transactions Database

Your firm maintains a `deals` table: deal_id, target_sector, deal_value, target_ebitda, deal_year,
acquirer_type. To find the median EV/EBITDA multiple paid for Indian cement sector targets in
strategic (not financial/PE) acquisitions over the past five years: SELECT
AVG(deal_value/target_ebitda) as avg_multiple FROM deals WHERE target_sector='Cement' AND
acquirer_type='Strategic' AND deal_year >= 2020; — and to see the full distribution rather than just
the average, you would pull all matching rows into pandas and compute .describe() for the full
quartile range, since a single average across a small deal sample can be highly misleading.

FROM RAMAN'S DESK

A regression on deal premiums against target characteristics — control premium paid versus target's
pre-deal growth rate, leverage, and strategic rationale category — is one of the more sophisticated
pieces of analysis I would ask a strong associate to run when building a takeover defense or an offer
recommendation. It moves the conversation from 'here are five deals we found' to 'here is the
statistically expected premium given this target's specific characteristics, and here is why our target
differs from that expectation.'

3.5 Football Field and Statistical Triangulation

No single valuation method is definitive. The professional standard — the 'football field' chart — presents the
range from each method (DCF, trading comps, precedent transactions, and where relevant, asset-based or
sum-of-the-parts) side by side, allowing the Investment Committee or client to see where methods agree and
where they diverge, and to weight their judgment accordingly.

Method Typical Range Driver When It's Most Reliable

DCF WACC and terminal growth Stable, predictable cash flow businesses with
assumptions limited comp availability
Trading Comps Public market sentiment and Sectors with a liquid, genuinely comparable
comp set selection public peer set
Precedent Transactions Control premiums, deal-specific M&A context specifically — less relevant for
strategic rationale minority stake valuation
Sum-of-the-Parts Accuracy of segment-level data Diversified conglomerates where a single
and segment-specific multiples blended multiple misrepresents the business

For Academic Use Only | Page 12

XLRI Jamshedpur | Data Analytics for Finance | Raman

3.6 Case Lab — Valuing an IPO-Bound Indian Company

CASE LAB — Valuing an IPO-Bound Indian Company
You are the lead analyst on a pre-IPO valuation for a mid-sized Indian specialty chemicals
manufacturer. You are given: five years of standardized financials, a list of fifteen potential public
comparables (with trading data), and eight precedent private placement transactions in the sector
from the past three years. Your deliverable: (1) Build a clean comp set from the fifteen candidates,
screening out any that fail basic size/growth/margin similarity tests — justify every exclusion. (2) Build
a base-case DCF and a WACC/terminal-growth sensitivity table. (3) Compute the implied multiple
range from the precedent transactions. (4) Present a football field chart triangulating all three
methods, and recommend an IPO price range with a one-paragraph rationale — the same deliverable
I would have expected from a VP-level banker before presenting to a client's board.

ANALYST'S CAUTION: The most common valuation error I have seen in twenty years of reviewing junior work
is false precision — presenting a DCF output to the exact rupee, or a single 'right' EV/EBITDA multiple, when
the underlying assumptions justify only a defensible range. A client or Investment Committee that senses
false precision will trust your work less, not more.

Module 3 — Self-Assessment
Q# Question
1 Explain the difference between a static and a data-fed valuation model, and describe one specific risk that
a data-fed approach mitigates.
2 You are building a comp set of 40 companies. Describe your process for identifying and handling statistical
outliers, and explain why silently deleting inconvenient comparables is a serious analytical error.
3 Design a two-variable Data Table sensitivity analysis for a DCF model — specify your two variables, your
range for each, and explain what the resulting grid tells an Investment Committee that a single base-case
number does not.
4 What is the conceptual difference between a Data Table sensitivity analysis and a Monte Carlo
simulation? When would you use one over the other?
5 Why might a precedent transaction multiple be systematically higher than a trading comparable multiple
for the same sector? What does this imply about which method to use for a minority stake valuation
versus a full acquisition?

For Academic Use Only | Page 13

XLRI Jamshedpur | Data Analytics for Finance | Raman

MODULE 4: MARKET & EQUITY ANALYTICS

4.1 Stock Price and Returns Analytics

Every equity research report ultimately rests on a foundation of return and risk statistics computed from raw
price data. Getting these calculations right — and understanding what they do and do not tell you — is
foundational to every other technique in this module.

4.1.1 Core Calculations

Metric Formula What It Tells You
Daily/Periodic Return (Pt - Pt-1) / Pt-1 Period-over-period price change
Annualized Volatility Std. Dev. of daily returns × √252 Risk measure — how much the stock's
price typically moves
Maximum Drawdown Largest peak-to-trough decline over a Worst-case historical loss an investor
period would have experienced
Rolling Beta Cov(stock, index) / Var(index), Time-varying sensitivity to market
computed over a rolling window moves — beta is not static

WORKED EXAMPLE — Rolling Beta in Python

Given daily returns for a stock and the Nifty 50 index in a DataFrame with columns stock_ret and
index_ret, a 60-day rolling beta is computed as: rolling_cov =
df['stock_ret'].rolling(60).cov(df['index_ret']); rolling_var = df['index_ret'].rolling(60).var();
df['rolling_beta'] = rolling_cov / rolling_var. Plotting this over several years often reveals that a
company's beta shifts materially around major events — a change in business mix, a leverage change,
or a macro regime shift — information a single static beta figure from a data vendor completely hides.

4.2 Factor Analytics for Equity Research

Institutional equity research increasingly frames stock selection through systematic factors rather than pure
narrative. Understanding how to construct and test these factors — even simply — makes you a far more
credible voice in any research or portfolio discussion.

Factor Typical Construction Investment Logic

Value Low P/E, low EV/EBITDA, or high Cheaper stocks relative to fundamentals tend to
earnings yield relative to peers outperform over long horizons
Momentum Strong price performance over the Stocks that have performed well tend to continue
trailing 6-12 months for a period (with caveats)
Quality High ROE, low leverage, stable Financially strong companies tend to be more
earnings growth resilient across cycles
Low Volatility Below-average historical price Lower-risk stocks have historically delivered
volatility surprisingly competitive risk-adjusted returns

FROM RAMAN'S DESK

At an equity research desk I advised, we built a simple quality-factor screen for Indian mid-caps — ROE

For Academic Use Only | Page 14

XLRI Jamshedpur | Data Analytics for Finance | Raman

above sector median, Debt/Equity below sector median, and 3-year earnings growth positive in every
year. It was not sophisticated. It was four lines of pandas code. But run consistently across every
quarterly universe update, it kept the desk out of at least two companies that later ran into serious
governance trouble. The lesson: a simple, disciplined, systematically-applied factor screen often
outperforms an ad hoc, purely narrative-driven stock picking process — not because it is smarter, but
because it is consistent.

4.3 Event Study Methodology

An event study measures whether a stock's return around a specific announcement — an earnings release, an
M&A announcement, a regulatory action — was abnormal relative to what the market as a whole was doing at
the same time. This is standard methodology in equity research to assess whether the market has already
'priced in' an event, or reacted with genuine surprise.

4.3.1 The Abnormal Return Calculation

Step Calculation
1. Estimate expected return Using a market model: Expected Return = alpha + beta × Market Return,
estimated over a clean pre-event window
2. Compute actual return The stock's actual return on the event date (and surrounding days)
3. Compute abnormal AR = Actual Return − Expected Return
return (AR)
4. Compute cumulative Sum of AR across the event window (e.g., day -1 to day +3) to capture the full
abnormal return (CAR) reaction, including any drift

WORKED EXAMPLE — Event Study — Earnings Surprise Reaction

A company's beta (estimated over the prior 120 trading days) is 1.1. On the day it reports a large
positive earnings surprise, the Nifty 50 index returns +0.5%. Expected return for the stock, per the
market model, is roughly 1.1 × 0.5% = 0.55%. The stock's actual return that day is +6.2%. Abnormal
Return = 6.2% − 0.55% = +5.65%. If the abnormal return over the following three trading days remains
meaningfully positive (rather than reverting), this suggests the market took time to fully digest the
surprise — a pattern known as post-earnings-announcement drift, and one that systematic equity
strategies have historically tried to exploit.

4.4 Sentiment and Alternative Data in Equity Research

Text — earnings call transcripts, management commentary, news flow — contains information that structured
financial data does not capture directly. You do not need advanced NLP expertise to extract real value here; a
disciplined, simple approach goes a long way.

Technique Practical Application

Keyword and tone scoring Counting positive vs. negative words (using finance-specific dictionaries like
Loughran-McDonald) in earnings call transcripts, tracked over time
Management language Comparing this quarter's MD&A language to prior quarters — a sudden shift in
change detection hedging language ('we believe' becoming 'we expect') can be a meaningful signal
Analyst Q&A tone analysis The tone and persistence of analyst follow-up questions on earnings calls often
signals market skepticism before it shows in the stock price

For Academic Use Only | Page 15

XLRI Jamshedpur | Data Analytics for Finance | Raman

News flow volume and Tracking the volume and average sentiment of news mentions can act as an early
sentiment proxy for shifting market perception

ANALYST'S CAUTION: Sentiment analysis on financial text is directionally useful, not a precision instrument.
Words like 'liability' or 'exposure' score as negative in general-purpose sentiment dictionaries but are entirely
neutral, routine terms in a financial filing. Always use finance-specific sentiment dictionaries, and always treat
sentiment scores as one input among several — never as a standalone signal.

4.5 Case Lab — Did the Market Overreact?

CASE LAB — Did the Market Overreact?
You are given daily price data for a real Indian company around a significant corporate action — an
unexpected regulatory penalty, a large M&A announcement, or a major earnings surprise (instructor
will assign the specific event) — along with Nifty 50 index data over the same period. Your task: (1)
Estimate the stock's beta over a clean 120-day pre-event window. (2) Compute daily abnormal returns
for a window spanning five days before to ten days after the event. (3) Plot the cumulative abnormal
return and identify whether the market's reaction appears to have been immediate and complete,
delayed (drift), or an overreaction later partially reversed. (4) Write a one-page research note stating
your interpretation and, critically, what an equity analyst covering this stock should have told clients
to do in the days immediately following the event.

Module 4 — Self-Assessment
Q# Question
1 Explain why a rolling beta calculation can reveal information that a single static beta figure hides. Give a
business scenario where you would expect beta to shift materially.
2 Design a simple two-factorcreen (choose any two factors from Value, Momentum, Quality, Low Volatility)
for an Indian mid-cap universe, and explain your rationale for combining these two specifically.
3 Walk through the four steps of an event study calculation for a hypothetical stock with beta 0.9, where
the market returns +1% and the stock returns +8% on the event date.
4 What is post-earnings-announcement drift, and how might a systematic strategy attempt to exploit it?
5 Why must sentiment analysis on earnings call transcripts use finance-specific dictionaries rather than
general-purpose sentiment tools? Give one example word that illustrates the problem.

For Academic Use Only | Page 16

XLRI Jamshedpur | Data Analytics for Finance | Raman

MODULE 5: CREDIT & RISK ANALYTICS

5.1 Credit Scoring Fundamentals

A credit score is, at its analytical core, a probability estimate — the probability that a borrower defaults within
a defined horizon, given their observable characteristics. Logistic regression remains the workhorse model
across most Indian banks and NBFCs for exactly this task, because it is transparent, auditable, and defensible
to a regulator in a way that more opaque models are not.

5.1.1 The Logistic Regression Framework for Default Prediction

Unlike linear regression, which predicts a continuous number, logistic regression predicts a probability
bounded between 0 and 1 — exactly the property needed for a default probability.
Concept Explanation
Dependent Variable Binary: 1 if the borrower defaulted within the observation window, 0 if not
Independent Variables Financial ratios: leverage, interest coverage, current ratio, profitability trend, and
often qualitative factors like industry and management quality scores
Output A probability of default (PD), typically between 0 and 1, for each borrower
Model Validation Out-of-sample testing on borrowers the model has never seen, using metrics like
the ROC-AUC curve and the Gini coefficient

WORKED EXAMPLE — Building a Simple Default Model (conceptual, Python/scikit-learn)

Given a DataFrame `df` with columns leverage, interest_coverage, current_ratio, roa, defaulted (0/1)
for a historical sample of borrowers: from sklearn.linear_model import LogisticRegression; X =
df[['leverage','interest_coverage','current_ratio','roa']]; y = df['defaulted']; model =
LogisticRegression().fit(X, y). The model's coefficients tell you the direction and relative importance of
each factor — typically, leverage and low interest coverage carry positive coefficients (higher default
probability), while ROA and current ratio carry negative coefficients. For a new borrower,
model.predict_proba() returns their estimated probability of default, which can then feed directly into
a credit approval or pricing decision.

ANALYST'S CAUTION: A model trained on data from a benign credit cycle will systematically underestimate
default risk once the cycle turns — this is precisely the failure mode I described in Module 1 regarding the
2008 crisis. Any credit model must be stress-tested against a downturn scenario, not just validated on
historical data drawn from stable conditions.

5.2 Portfolio Risk Analytics

A single loan's risk is manageable in isolation. A portfolio of thousands of loans requires aggregate risk
measures — how much could the whole book lose, and under what conditions.

For Academic Use Only | Page 17

XLRI Jamshedpur | Data Analytics for Finance | Raman

5.2.1 Value at Risk (VaR)

VaR answers a specific question: at a given confidence level, what is the maximum loss the portfolio is
expected NOT to exceed over a defined time horizon? For example, a 1-day 95% VaR of Rs. 10 crore means
there is a 95% probability the portfolio will not lose more than Rs. 10 crore in a single day.
VaR Method Approach Strength Weakness
Historical Simulation Apply actual historical return No distributional Assumes the future
distribution to the current assumption required resembles the
portfolio historical sample
period
Parametric (Variance- Assumes returns are normally Fast, simple, widely Underestimates tail
Covariance) distributed; VaR computed from understood risk if returns are not
mean, std. dev. actually normal (they
rarely are)
Monte Carlo Simulation Simulates thousands of possible Flexible, can Computationally
future portfolio paths using incorporate complex heavier; still
assumed distributions correlations and non- dependent on the
normal distributions quality of the
assumed
distributions

5.2.2 Stress Testing

Where VaR tells you the likely range of outcomes under normal conditions, stress testing asks a deliberately
extreme question: what happens to this portfolio if a specific severe scenario materializes — a sharp interest
rate shock, a sector-wide downturn, a sudden currency devaluation?

WORKED EXAMPLE — Stress Testing a Loan Portfolio

A bank's real estate loan book totals Rs. 5,000 crore. A stress scenario assumes: (a) property values
decline 25%, (b) interest rates rise 200 basis points, and (c) borrower incomes in the affected segment
fall 15%. Applying these shocks to the loan-level data — recalculating loan-to-value ratios, debt
service coverage ratios, and re-estimating default probabilities under the stressed inputs using the
logistic regression model from Section 5.1 — produces a stressed expected loss figure. If this stressed
loss exceeds the bank's provisioning buffer for this segment, that is a Board-level flag, not merely an
analyst's note.

5.3 Early Warning Systems

The most valuable work a credit risk analyst does is not scoring a new loan application — it is catching a
performing loan before it becomes a non-performing one. This requires ongoing, systematic monitoring, not
periodic manual review.

Trigger Type Example Typical Threshold

Covenant-based Interest coverage ratio falls below a As specified in the loan agreement —
contractually specified minimum often 1.5x to 2.0x
Rating migration Internal or external credit rating Firm-specific, but rapid multi-notch
downgraded by two or more notches within downgrades are universally significant
a year
Behavioral Increased utilization of a working capital Utilization consistently above 90% of
facility close to its sanctioned limit sanctioned limit

For Academic Use Only | Page 18

XLRI Jamshedpur | Data Analytics for Finance | Raman

Statistical Model-estimated probability of default rises Often a relative trigger — e.g., PD

above a defined threshold from the prior increase of 50%+ quarter-on-quarter
quarter
Qualitative/News-based Adverse news flow, auditor qualification, Judgment-based, but should feed into
management exits, related-party the same monitoring dashboard
transaction flags

5.4 Bank and NBFC-Specific Analytics — NPA and Asset Quality Review
Indian banking and NBFC analysis carries its own specific vocabulary and regulatory framework that a
generalist credit model does not automatically capture. Understanding NPA classification and its analytical
implications is essential for anyone covering the sector, whether as a lender, an equity analyst, or a regulator.

Concept Definition Analytical Significance

Gross NPA (GNPA) Total loans classified as non- Headline asset quality indicator; trend over
performing (90+ days overdue), time is the key signal
before provisions
Net NPA (NNPA) Gross NPA minus provisions already Reflects the bank's true remaining exposure
made against those loans after accounting for provisioning
Provision Coverage Provisions held / Gross NPAs A low PCR relative to peers suggests under-
Ratio (PCR) provisioning and future earnings risk
Slippage Ratio New NPAs added during a period / A forward-looking flow measure — often
Standard advances at the start of more informative than the stock measure
the period (GNPA%) alone
Restructured Assets Loans modified in terms due to Often a leading indicator — a rising
borrower stress but not yet restructured book frequently precedes a
classified NPA rising NPA book

FROM RAMAN'S DESK

When I advised on distressed asset situations for Indian banks, the single most useful chart I could
produce for a Board was not the reported GNPA ratio — everyone already had that. It was the
slippage ratio trend alongside the restructured book trend, plotted together over eight quarters. A
rising restructured book with a stable slippage ratio often means trouble is being deferred, not
resolved. That combination, more than any single reported ratio, told the real story of where the credit
cycle in that book was actually heading.

5.5 Case Lab — Building a Credit Risk Dashboard

CASE LAB — Building a Credit Risk Dashboard for a Mid-Market Lending Book
You are given loan-level data for a mid-market NBFC's 2,000-loan book: borrower financial ratios,
days-past-due history, sector classification, and loan terms. Your deliverable: (1) Build a logistic
regression default model using a training subset of the data, and validate it on a held-out test subset.
(2) Compute portfolio-level expected loss under current conditions and under a stress scenario
(specified by the instructor — e.g., a sector-specific downturn). (3) Design and apply at least three
early warning triggers (covenant, behavioral, and statistical) to flag the highest-risk accounts in the
current book. (4) Build a one-page dashboard — the kind I would expect to review weekly as a credit
committee member — summarizing portfolio health, the top ten highest-risk accounts, and your top
three recommended actions.

For Academic Use Only | Page 19

XLRI Jamshedpur | Data Analytics for Finance | Raman

Module 5 — Self-Assessment
Q# Question
1 Explain why a credit model trained only on data from a benign credit cycle poses a specific risk, and
describe how stress testing addresses this.
2 Compare parametric and historical simulation approaches to VaR. Under what market conditions would
you expect the parametric approach to significantly underestimate true risk?
3 Design three early warning triggers (one covenant-based, one behavioral, one statistical) for a corporate
lending portfolio, specifying the exact threshold you would use for each.
4 Explain the difference between Gross NPA and Net NPA, and explain why a bank with a low Provision
Coverage Ratio poses additional analytical concern beyond its headline GNPA%.
5 Why might a rising restructured assets book, even with a stable slippage ratio, still be considered a
warning sign by an experienced credit analyst?

For Academic Use Only | Page 20

XLRI Jamshedpur | Data Analytics for Finance | Raman

MODULE 6: M&A AND DEAL ANALYTICS

6.1 Deal Sourcing and Screening Analytics

Before any deal reaches a bank's or PE fund's Investment Committee, someone had to find it — usually from a
universe of hundreds or thousands of candidates. Systematic screening, rather than purely relationship-driven
deal flow, has become the standard first filter at every serious institution I worked with.

6.1.1 Building a Multi-Criteria Screen

Screening Criterion Typical Filter Purpose
Size Revenue or EBITDA within the fund's Eliminates candidates too small or too large
or acquirer's target deal-size band to be actionable
Growth profile Revenue CAGR above a defined Filters for businesses matching the strategic
threshold over the trailing 3 years growth thesis
Profitability EBITDA margin above a sector- Filters out structurally weak businesses
relative threshold regardless of growth
Ownership structure Promoter shareholding pattern, Flags candidates with a plausible reason to
presence of existing PE investors, transact (e.g., succession gap, existing
succession situation investor seeking exit)
Leverage Debt/EBITDA below a threshold, or Aligns with the specific mandate — growth
conversely, distressed leverage for equity vs. distressed/special situations
special-situations mandates

WORKED EXAMPLE — Screening a 500-Company Universe in pandas

A PE fund's mandate is mid-market Indian consumer companies with Rs. 50-300 crore revenue,
EBITDA margin above 15%, and 3-year revenue CAGR above 12%, with no existing institutional PE
investor on the cap table. Starting from a 500-company screened universe DataFrame `df`: shortlist =
df[(df['revenue'].between(50,300)) & (df['ebitda_margin']>0.15) & (df['revenue_cagr_3y']>0.12) &
(df['existing_pe_investor']==False)]. In one line, a 500-company universe is reduced to perhaps fifteen
to twenty-five genuinely actionable candidates — the starting point for the actual relationship-
building and diligence work that follows.

6.2 Synergy Quantification

Every M&A pitch deck claims synergies. The discipline that separates a credible deal thesis from an optimistic
one is whether those synergies are quantified from real operational data, bottoms-up, rather than asserted
top-down as a convenient percentage of combined revenue.

6.2.1 Bottoms-Up Synergy Categories

Synergy Type How It Is Quantified Realism Check
Procurement / Cost Combined purchasing volume applied Cross-check against what similar-scale
Synergies against a vendor-negotiated discount companies in the sector actually achieve
curve, category by category — is the assumed discount curve
realistic?

For Academic Use Only | Page 21

XLRI Jamshedpur | Data Analytics for Finance | Raman

Headcount / Overhead Duplicate functions (finance, HR, certain Factor in genuine one-time severance
Synergies back-office roles) mapped and and transition costs — rarely fully
consolidated headcount estimated role modeled by optimistic bankers
by role
Revenue Synergies Acquirer's customer base × estimated Historically the most overestimated
(Cross-Sell) attach rate for target's product × target's synergy category — apply a significant
average revenue per customer haircut and a longer realization timeline
Facility / Capex Combined facility utilization analysis Requires genuine operational data, not
Synergies identifying redundant capacity that can just a percentage assumption
be consolidated or avoided

FROM RAMAN'S DESK

In my consulting years, I built more synergy models than I can count, and the single most reliable
predictor of whether announced synergies would actually be realized was whether the underlying
model was bottoms-up (role-by-role, vendor-category-by-vendor-category) or top-down (a flat 5% of
combined revenue). Boards and clients increasingly ask for the bottoms-up version now, because the
market has become skeptical of round-number synergy claims — and rightly so. If you cannot show
me the specific roles, the specific vendor contracts, or the specific customer overlap driving your
synergy number, I do not believe the number.

6.3 Accretion/Dilution Analytics

For a public acquirer, whether a deal is EPS-accretive or dilutive in the near term is one of the first questions
the market and the Board will ask — even though it is far from the only consideration that should drive a deal
decision. Building this analysis with proper sensitivity to financing structure is a core banking skill.

6.3.1 The Core Mechanics

Pro forma combined EPS depends critically on the financing mix — cash, debt, or stock — and each has a
different accretion/dilution effect.
Financing Method Accretion/Dilution Driver
All-Cash (using balance Accretive if the target's earnings yield exceeds the foregone interest income on
sheet cash) the cash used
All-Debt Accretive if the target's earnings yield exceeds the after-tax cost of the new debt
All-Stock Accretive only if the acquirer's P/E is higher than the target's implied acquisition
P/E (the 'bigger P/E buys smaller P/E' rule of thumb)

WORKED EXAMPLE — Accretion/Dilution Sensitivity Table

An acquirer trading at 22x P/E proposes to buy a target at an implied 16x P/E, funded via a mix of
stock and debt. Building a two-variable sensitivity table across the percentage of stock-vs-debt
financing (0% to 100% stock, in 20% increments) and the assumed synergy realization level (0%, 50%,
100% of modeled synergies) produces a grid showing pro forma EPS impact under each combination.
This is precisely the exhibit a banker presents to a CFO deciding on deal structure — it converts an
abstract financing choice into a concrete, quantified EPS outcome across a realistic range of scenarios.

For Academic Use Only | Page 22

XLRI Jamshedpur | Data Analytics for Finance | Raman

6.4 Post-Merger Integration Tracking

The deal model does not end at signing. The most operationally mature acquirers — and the best-run PE
portfolio companies — track actual post-close performance against the original deal-model assumptions with
the same analytical discipline used to build the original model.

Tracked KPI Purpose

Realized synergies vs. modeled synergies, by Direct accountability against the deal thesis presented to
category and by quarter the Board or Investment Committee
Customer retention rate (acquirer and target bases) Early warning if the deal is disrupting either business's
core customer relationships
Key employee retention Loss of critical talent frequently destroys more value
than any synergy shortfall
Combined entity margin trajectory vs. plan The ultimate scoreboard — is the combined business
actually performing as modeled

6.5 Case Lab — Screen, Shortlist, Structure

CASE LAB — Screen, Shortlist, Structure
You are working a live-style mandate: a mid-market PE fund wants to identify and structure an initial
offer for an add-on acquisition to an existing portfolio company in the specialty logistics sector. You
are given a 200-company screened universe with financial data, and detailed financials for the
acquirer (the existing portfolio company). Your deliverable: (1) Apply a multi-criteria screen to
shortlist five to eight actionable targets, justifying your screening logic. (2) For your top candidate,
build a bottoms-up synergy estimate (cost and revenue) using the operational data provided. (3) Build
an accretion/dilution sensitivity table across financing structure and synergy realization assumptions.
(4) Recommend a deal structure and headline valuation range, and identify the three KPIs you would
track most closely in the first year post-close.

ANALYST'S CAUTION: Revenue synergies are the most commonly overestimated line item in any deal model I
have reviewed in thirty-eight years. Cost synergies, drawn from concrete headcount and vendor data, are
difficult but achievable to forecast reliably. Revenue synergies depend on customer behavior, competitive
response, and execution quality that no model can fully anticipate. Apply a meaningfully larger haircut and a
longer realization timeline to revenue synergies than to cost synergies — always.

Module 6 — Self-Assessment
Q# Question
1 Design a five-criterion screen for a strategic acquirer in the Indian packaged foods sector looking for a
bolt-on acquisition, specifying the exact filter for each criterion.
2 Explain why a bottoms-up synergy model is generally more credible than a top-down percentage-of-
revenue synergy assumption. What specific data would you need to build a bottoms-up cost synergy
estimate?
3 An acquirer trading at 18x P/E is considering an all-stock acquisition of a target with an implied 24x
acquisition P/E. Will this deal be accretive or dilutive to the acquirer's EPS, all else equal? Explain the
mechanism.
4 Why should revenue synergies typically receive a larger haircut and longer realization timeline than cost
synergies in a deal model?
5 List four KPIs you would track in the first year after closing an acquisition to assess whether the deal is

For Academic Use Only | Page 23

XLRI Jamshedpur | Data Analytics for Finance | Raman

performing against its original thesis, and explain what a shortfall in each would signal.

For Academic Use Only | Page 24

XLRI Jamshedpur | Data Analytics for Finance | Raman

MODULE 7: PREDICTIVE & MACHINE LEARNING APPLICATIONS IN

FINANCE

MODULE 7: PREDICTIVE & MACHINE LEARNING APPLICATIONS IN

FINANCE
Learning Objectives

By the end of this module you should be able to: (1) Distinguish where machine learning genuinely
adds value in finance from where it is hype. (2) Build regularized regression models for driver-based
financial forecasting. (3) Build classification models for fraud, default, and churn prediction. (4) Apply
clustering techniques for portfolio and client segmentation. (5) Extract structured signal from
unstructured financial text at scale. (6) Build a simple ML-based target screening model for deal
sourcing.

7.1 Where ML Genuinely Adds Value in Finance — And Where It Doesn't

I want to be direct with you, because this is a question I have been asked in nearly every boardroom I have sat
in over the past decade: does machine learning actually make better financial decisions, or does it just look
impressive in a pitch deck? The honest answer is that it depends entirely on the use case, and a large fraction
of the 'AI-powered' claims I have encountered in vendor pitches did not survive serious scrutiny.

ML Genuinely Adds Value When... ML Is Often Oversold When...

The problem has abundant, clean historical data with The 'training data' is thin, unrepresentative, or drawn
a clear, measurable outcome (default/no default, only from benign conditions (see Module 1's 2008
fraud/no fraud) warning)

The relationships between variables are genuinely A simple, transparent regression would perform
complex and non-linear, beyond what a human nearly as well and is far more explainable to a
analyst or simple regression can capture regulator or client

The task is a single, high-stakes, judgment-heavy

The task is high-volume and repetitive — screening
decision — a major M&A call, a strategic pivot —
thousands of loan applications or thousands of
where human judgment and qualitative context
potential deal targets
dominate

The model's output feeds into a human decision- The model is a 'black box' that cannot explain its own
maker who can sanity-check it, not one that acts fully reasoning, in a regulated context requiring
autonomously on high-stakes capital explainability

FROM RAMAN'S DESK

A fintech once pitched my former firm an 'AI-powered' credit model that they claimed outperformed
traditional scoring by a wide margin. When we asked to see the model's performance specifically

For Academic Use Only | Page 25

XLRI Jamshedpur | Data Analytics for Finance | Raman

during a stressed period — not just the benign years it was trained and tested on — the answer was
an uncomfortable silence. That silence told us everything. I am not anti-machine-learning; some of
the most useful work in this module genuinely depends on it. I am anti-blind-faith. Every model, ML
or otherwise, must answer the question: how do you know this works, and specifically, how do you
know it works when conditions turn adverse?

7.2 Regression-Based Forecasting for FP&A and Revenue Prediction

Multiple linear regression, extended with regularization techniques, remains the most widely used and most
defensible predictive technique in corporate FP&A and revenue forecasting — precisely because it is
interpretable.

7.2.1 Ridge and LASSO Regularization

When you have many potential predictor variables — macro indicators, marketing spend, pricing, competitor
actions, seasonality — ordinary regression can overfit, especially with a limited number of historical periods.
Regularization techniques penalize model complexity to produce a more robust, generalizable forecast.
Technique What It Does Practical Effect

Reduces overfitting while keeping all

Shrinks all coefficients toward zero, variables in the model — useful when you
Ridge Regression (L2)
proportionally believe most variables genuinely matter
somewhat

Performs automatic variable selection —

Can shrink some coefficients exactly
LASSO Regression (L1) useful when you suspect many candidate
to zero
variables are actually irrelevant

A practical middle ground, often the default

Elastic Net Combines Ridge and LASSO penalties choice when uncertain which pure approach
fits better

WORKED EXAMPLE — Driver-Based Revenue Forecast with LASSO

An FP&A team wants to forecast next-quarter revenue for a retail chain using twelve candidate
predictors: prior quarter revenue, footfall, marketing spend, average ticket size, competitor store
openings, regional GDP growth, festive season indicator, and five others. With only 40 historical
quarters of data, ordinary regression on twelve variables risks severe overfitting. Applying LASSO —
from sklearn.linear_model import Lasso; model = Lasso(alpha=0.1).fit(X, y) — the model may
automatically shrink six of the twelve coefficients to exactly zero, leaving a leaner, more robust six-
variable forecast model that generalizes better to new quarters than the full twelve-variable model
would.

For Academic Use Only | Page 26

XLRI Jamshedpur | Data Analytics for Finance | Raman

7.3 Classification Models for Finance

Beyond the credit default classification covered in Module 5, classification techniques extend naturally to
fraud detection and client churn prediction — two problems every large financial institution invests heavily in.

7.3.1 Model Choices and Trade-offs

Model Strength Weakness Best Finance Use Case

Highly interpretable; Cannot capture complex Credit scoring where

Logistic Regression coefficients have clear non-linear interactions explainability is
meaning; regulator-friendly between variables regulatorily required

Initial exploratory fraud

Prone to overfitting if not
Easy to visualize and explain rule-building; explaining a
Decision Trees carefully pruned; can be
as a set of business rules decision to non-technical
unstable
stakeholders

Handles non-linear
Less directly interpretable Fraud detection where
relationships well; robust to
Random Forest than a single tree or logistic accuracy is prioritized
overfitting; strong out-of-
regression over full explainability
the-box performance

High-stakes fraud and

Requires more careful
Often the highest raw churn models where
Gradient Boosting tuning; higher risk of
predictive accuracy among marginal accuracy gains
(XGBoost) overfitting if not validated
classical ML methods have large financial
properly
impact

WORKED EXAMPLE — Fraud Detection Model Evaluation

A bank builds a Random Forest fraud detection model on transaction data. On the test set, it
correctly flags 850 of 1,000 actual fraudulent transactions (85% recall) but also flags 4,000 legitimate
transactions as fraudulent out of 500,000 (a low false-positive rate, but in absolute terms, 4,000
customers experiencing unnecessary friction). This is the central trade-off in every fraud model:
recall (catching real fraud) versus precision (not annoying legitimate customers). The 'right' balance is
a business decision, not purely a technical one — I have sat in exactly this conversation between a
risk team wanting maximum recall and a customer experience team wanting maximum precision,
and the resolution always depended on the specific cost of each type of error to that particular
institution.

7.4 Clustering for Portfolio and Client Segmentation

Unsupervised learning — where there is no labeled 'right answer' to predict, only patterns to discover — is
particularly useful for segmentation tasks in wealth management, portfolio construction, and peer grouping
that goes beyond simplistic sector classification.

For Academic Use Only | Page 27

XLRI Jamshedpur | Data Analytics for Finance | Raman

7.4.1 K-Means Clustering

K-means groups observations into a specified number of clusters based on similarity across chosen
dimensions, without any predefined labels.

WORKED EXAMPLE — Investor Segmentation for a Wealth Management Practice

A wealth manager has client data including portfolio size, risk tolerance score, average holding
period, trading frequency, and product mix (equity vs. debt vs. alternatives). Applying K-means
clustering (from [Link] import KMeans; KMeans(n_clusters=4).fit(X)) on standardized
versions of these variables might reveal four natural client segments: 'buy-and-hold conservative,'
'active trader aggressive,' 'goal-based systematic investor,' and 'alternatives-heavy sophisticated
investor' — segments the firm can then serve with genuinely differentiated product offerings and
communication strategies, rather than treating every client identically or segmenting purely by AUM
size, which often misses the real behavioral differences that matter for service design.

7.4.2 Clustering for Peer Grouping Beyond Sector Classification

Standard sector classifications (GICS, sector indices) are often too coarse for genuine peer comparison.
Clustering companies on actual financial and operational characteristics — margin structure, capital intensity,
growth profile, leverage — frequently produces more analytically meaningful peer groups than the official
sector label alone.

7.5 Text Analytics on Financial Documents

Annual reports, particularly the Management Discussion & Analysis (MD&A) section, contain risk disclosures
and forward-looking language that structured financial data never captures. Extracting signal from this text at
scale — across hundreds of companies, every year — is now a standard part of sophisticated equity research
and credit analysis.

Technique Financial Application

Tracking the frequency of terms like 'going concern,' 'covenant,' 'liquidity,' or

Keyword frequency
'litigation' across a company's annual reports year over year — a rising frequency
tracking over time
is a legitimate flag worth investigating

Academic research has repeatedly found that annual reports with deliberately
Readability and complexity
more complex, harder-to-read language are correlated with weaker subsequent
scoring
financial performance and, in some cases, earnings management

Named entity and topic Systematically identifying which specific risks (regulatory, litigation, competitive,
extraction currency) a company discusses, and how this risk-topic mix shifts year over year

FROM RAMAN'S DESK

At a Big Four advisory practice, we built a simple text-scoring tool that tracked the frequency of
hedging language — 'may,' 'could,' 'believe,' 'expect' as opposed to more definitive statements —
across three years of a client's MD&A disclosures ahead of a due diligence engagement. It was not

For Academic Use Only | Page 28

XLRI Jamshedpur | Data Analytics for Finance | Raman

sophisticated NLP by any modern standard. But the trend was unmistakable: hedging language had
increased steadily for two years before the engagement began. Combined with the ratio analytics
from Module 2, it painted a consistent picture that the standalone financial statements alone had
not fully revealed. Text is a legitimate data source. Treat it with the same rigor you would treat a
balance sheet.

7.6 Case Lab — Build a Target Screening ML Model

CASE LAB — Build a Target Screening ML Model for Deal Sourcing

You are building a screening tool for a VC/PE-style deal sourcing mandate. You are given a dataset of
300 companies with financial, operational, and text-derived features (including a simple sentiment
score derived from their public disclosures), along with a label indicating which of these companies
subsequently received institutional funding within two years (your training signal). Your deliverable:
(1) Build and compare at least two classification models (e.g., Logistic Regression and Random
Forest) to predict funding likelihood. (2) Evaluate both models using an appropriate metric beyond
simple accuracy (precision, recall, and ROC-AUC), and explain which model you would recommend
and why — including the explainability trade-off discussed in Section 7.3. (3) Apply your chosen
model to a fresh, unlabeled set of 50 companies and produce a ranked shortlist. (4) Present your
shortlist alongside a clear statement of the model's limitations — what it cannot tell you, and what
human judgment must still be applied before any of these companies are actually approached.

ANALYST'S CAUTION: The single most dangerous mistake in applying machine learning to finance is treating
a model's output as a conclusion rather than as one input into a human decision. Every model in this module
— from a simple logistic regression to a gradient-boosted classifier — is a tool for organizing and weighing
evidence at scale. It is not, and should never be presented as, a replacement for the judgment of the person
accountable for the capital being deployed.

Module 7 — Self-Assessment
Q# Question

Describe a specific finance use case where you believe a simple logistic regression would be preferable to
1
a more complex model like XGBoost, and explain why, referencing the explainability trade-off.

Explain the difference between Ridge and LASSO regularization, and describe a forecasting scenario
2
where you would specifically prefer LASSO's variable-selection property.

A fraud detection model has high recall but low precision. Explain what this means in plain business
3
terms, and describe the trade-off a bank's risk and customer experience teams would need to negotiate.

Design a K-means clustering approach (specify your input variables) to segment a bank's SME lending
4
clients into meaningful groups beyond simple loan-size bands.

Why is it dangerous to validate a credit or fraud model only on data drawn from a single, benign
5
economic period? How does this connect back to the Module 1 discussion of the 2008 crisis?

For Academic Use Only | Page 29

XLRI Jamshedpur | Data Analytics for Finance | Raman

MODULE 8: CAPSTONE — THE ANALYST'S DELIVERABLE

MODULE 8: CAPSTONE — THE ANALYST'S DELIVERABLE

Learning Objectives

By the end of this module you should be able to: (1) Structure a data-driven analysis into a client-
ready Investment Committee memo. (2) Build a clean, decision-focused dashboard following
professional visualization standards. (3) Integrate techniques from every prior module into a single,
coherent, defensible recommendation, and defend it live under questioning.

8.1 From Analysis to Investment Committee Memo

This is, in many ways, the most important module in the entire course — and the one most MBA programs
skip entirely. You can run the most sophisticated regression in the world, and if you cannot compress it into a
one-page memo that a Managing Director can read in ninety seconds and act on, the analysis has no
commercial value. I have watched brilliant analytical work die in a drawer because nobody could explain,
quickly enough, why it mattered.

8.1.1 The One-Page Memo Discipline

Section Length Content

1-2 sentences, at
State your conclusion first — buy, pass, flag for monitoring,
Recommendation the top, not buried
proceed with the deal
at the end

The two or three data-driven findings that most directly support

Why 3-4 bullet points
the recommendation

The single biggest reason you could be wrong — every credible

Key Risk 1-2 sentences
analyst names this explicitly

Attached, not in Full models, dashboards, and data tables live in an appendix for
Supporting Detail
the memo body anyone who wants to go deeper

FROM RAMAN'S DESK

At McKinsey, we had a phrase for this: 'the pyramid principle' — lead with the answer, then the
supporting logic, then the detail, in that order, always. In thirty-eight years, I have never once had a
client or an Investment Committee ask me to make a memo longer. I have been asked, more times
than I can count, to make one shorter. Train yourself now, as a student, to write the one-page
version first — before the twenty-page version — because the discipline of compressing your own
thinking into one page is what will make you promotable faster than any technical skill in this book.

For Academic Use Only | Page 30

XLRI Jamshedpur | Data Analytics for Finance | Raman

8.2 Dashboard and Visualization Standards

A dashboard exists to enable a decision, not to demonstrate how much data you have access to. The discipline
of good financial visualization is fundamentally a discipline of subtraction — removing everything that does
not directly serve the decision at hand.

Principle In Practice

Every chart should answer a single, clearly identifiable question — not attempt to
One chart, one question
show everything you know at once

Lead with the conclusion in A chart titled 'Receivables Days, 2021-2025' is weaker than one titled
the title 'Receivables Days Have Risen 40% Since Q3 2024'

3D effects, unnecessary gridlines, excessive color, and decorative elements

Remove chartjunk
actively reduce comprehension speed — remove them all

Consistent color logic If red means 'risk' or 'below threshold' on one page, it must mean the same thing
across a deck on every page — inconsistency erodes trust

A number in isolation is far less useful than the same number shown against a
Benchmark, always
peer average, a historical trend, or a target

8.3 Final Capstone Project

CASE LAB — The Capstone: A Full Investment or Credit Recommendation

Working individually or in a team of up to three, you will select (subject to instructor approval) one
of the following tracks, mirroring the exact structure of a real engagement I would have staffed a
junior team on: TRACK A — Equity Research: Full analysis of a listed Indian company, integrating
financial statement analytics (Module 2), valuation triangulation (Module 3), and market/event-study
analytics (Module 4), culminating in a buy/hold/sell recommendation. TRACK B — Credit
Recommendation: Full credit analysis of a mid-market borrower, integrating ratio and trend analytics
(Module 2) with a default probability model and early warning framework (Module 5), culminating in
an approve/decline/monitor recommendation with proposed terms. TRACK C — M&A/Investment
Thesis: Full target screening, valuation, and synergy analysis (Modules 3 and 6) for a hypothetical
acquisition or PE investment, culminating in a go/no-go recommendation with proposed structure.
Every track must include: a one-page Investment Committee memo (Section 8.1 discipline), a
supporting dashboard (Section 8.2 standards), and the underlying analytical model. You will present
and defend your recommendation live, in under seven minutes, followed by five minutes of
questioning — exactly the format of every Investment Committee I sat on for thirty-eight years.

FROM RAMAN'S DESK

I will ask you the same three questions I asked every analyst who ever presented to me: What is your
recommendation, in one sentence? What is the single strongest piece of evidence for it? And what is
the one thing that would make you change your mind? If you can answer all three without
hesitation, you have done the work properly — regardless of how sophisticated or simple the
underlying technique was. That is the standard this entire course has been building toward, session
by session, since Module 1.

For Academic Use Only | Page 31

XLRI Jamshedpur | Data Analytics for Finance | Raman

For Academic Use Only | Page 32

XLRI Jamshedpur | Data Analytics for Finance | Raman

APPENDIX A: MASTER TOOLKIT REFERENCE

Tool / Technique Module Core Use

Automated, refreshable data pipelines from messy

Power Query 1
source files

SQL — SELECT, JOIN, GROUP BY, Window Precise extraction and aggregation from structured
1
Functions databases

1,
Tabular data manipulation, ratio computation, filtering
pandas / NumPy through
at scale
out

Ratio Analytics
(Liquidity/Leverage/Profitability/Efficienc 2 Financial health assessment across peer universes
y)

Converting absolute figures into relative, decision-useful

Percentile Benchmarking 2
context

Two-variable sensitivity analysis for valuation

Data Tables (What-If Analysis) 3
assumptions

Full probability distribution of valuation outcomes

Monte Carlo Simulation 3
across many assumptions

Comparable Company / Precedent

3 Market-based valuation cross-checks
Transaction Analysis

Rolling Beta, Volatility, Drawdown 4 Risk and return characterization from raw price data

Factor Construction
4 Systematic equity screening and portfolio tilts
(Value/Momentum/Quality)

Event Study / Abnormal Returns 4 Measuring market reaction to corporate events

Logistic Regression for Default Prediction 5 Credit scoring and probability-of-default estimation

Portfolio-level risk quantification under normal and

Value at Risk (VaR) and Stress Testing 5
stressed conditions

Early Warning Triggers 5 Systematic monitoring for credit deterioration

Synergy Quantification (bottoms-up) 6 Credible cost and revenue synergy estimation in M&A

Accretion/Dilution Analysis 6 EPS impact assessment across financing structures

Ridge / LASSO Regression 7 Regularized, overfitting-resistant forecasting models

Classification Models (Logistic, Trees,

7 Fraud, default, and churn prediction
Random Forest, XGBoost)

Client and company segmentation beyond standard

K-Means Clustering 7
classifications

Signal extraction from MD&A, earnings calls, and

Financial Text Analytics 7
disclosures

One-Page IC Memo Structure 8 Compressing analysis into a decision-ready

For Academic Use Only | Page 33

XLRI Jamshedpur | Data Analytics for Finance | Raman

Tool / Technique Module Core Use

recommendation

For Academic Use Only | Page 34

XLRI Jamshedpur | Data Analytics for Finance | Raman

APPENDIX B: KEY TERMS GLOSSARY

Term Definition

The portion of a stock's return not explained by overall market movement,

Abnormal Return
used in event studies

Accretion/Dilution The impact of an acquisition on the acquirer's pro forma earnings per share

Comp Set (Comparable Company A group of publicly traded companies used as a market-based valuation
Set) benchmark

A visual comparison of valuation ranges from multiple methods (DCF,

Football Field Chart
comps, precedents) side by side

The principle that a model's output is only as reliable as the quality of its
GIGO (Garbage In, Garbage Out)
input data and assumptions

Regularization techniques that penalize model complexity to reduce

LASSO / Ridge Regression
overfitting

A technique using randomized repeated sampling to produce a full

Monte Carlo Simulation
distribution of possible outcomes

A loan on which the borrower has not made scheduled payments for a
NPA (Non-Performing Asset)
specified period (90+ days in India)

Valuation benchmarking using multiples paid in past comparable M&A

Precedent Transaction Analysis
deals

A model-estimated probability that a borrower will default within a

Probability of Default (PD)
defined time horizon

Provisions held against NPAs, divided by Gross NPAs — a measure of a

Provision Coverage Ratio (PCR)
lender's provisioning conservatism

A family of techniques (Ridge, LASSO) that constrain model complexity to

Regularization
improve out-of-sample generalization

New NPAs added in a period, divided by standard advances at the start of

Slippage Ratio
the period — a forward-looking asset quality flow measure

Standardization (Financial The process of remapping different companies' financial statements onto a
Statements) common, comparable chart of accounts

The additional value created by combining two businesses, beyond what

Synergy (Cost / Revenue)
each could achieve independently

The maximum expected loss on a portfolio, at a given confidence level,

Value at Risk (VaR)
over a defined time horizon

For Academic Use Only | Page 35

XLRI Jamshedpur | Data Analytics for Finance | Raman

APPENDIX C: RECOMMENDED PRACTICE RESOURCES

Resource Use

[Link] Quick Indian company financial screening for coursework and self-practice

NSE/BSE Bhavcopy Free daily price and volume data for building your own return/volatility
([Link] / [Link]) analytics

Public datasets for practicing classification and regression techniques

Kaggle Financial Datasets
outside class

Company Annual Reports (via Primary-source practice for financial statement standardization and MD&A
BSE/NSE filings) text analysis

RBI Database on Indian Economy Macro and banking sector data for credit risk and macro-driven analytics
(DBIE) practice

Python Libraries: pandas, NumPy,

The complete technical toolkit used throughout this course
scikit-learn, matplotlib

For Academic Use Only | Page 36

XLRI Jamshedpur | Data Analytics for Finance | Raman

A CLOSING NOTE FROM RAMAN

Thirty-eight years ago, I started my career as a junior analyst who could barely build a pivot table. I did not
become useful to the people who trusted me with real capital because I learned every technique in this book
— I did not know most of them yet. I became useful because I developed, slowly, over many mistakes, the
judgment to know when a number was telling me the truth and when it was hiding something.

Every technique in this book — from a simple SUMIFS formula to a gradient-boosted classification model — is
in service of that same judgment. Tools will keep changing. When I started, we did not have Bloomberg
terminals on every desk, let alone Python. In another decade, some of what I have taught you here will look as
dated as the tools I started with. What will not change is the discipline underneath: trace every number to its
source, understand what a model can and cannot tell you, and never present a conclusion you cannot defend
in one sentence to someone who was not in the room when you built it.

I told the Director I agreed to teach this course out of passion for the subject. That was true. But I will tell you,
now that we have reached the end of this book together, the fuller reason: I spent thirty-eight years being
taught, in expensive and sometimes painful ways, by markets, by clients, and by a handful of extraordinary
mentors of my own. Writing this book, and teaching this class, is the only way I know to pass that forward
before it is too late to matter.

Go build something worth defending.

Raman
XLRI Jamshedpur

— END OF TEXTBOOK —

XLRI Jamshedpur | Data Analytics for Finance | MBA Finance Elective

First Edition, 2025 | For Academic Use Only

For Academic Use Only | Page 37