FORECASTING — Complete Study Notes
(Stevenson, Chapter 3 — up to Exponential Smoothing)
1. WHAT IS FORECASTING?
Forecast = a statement about the future value of a variable of interest
(e.g., demand).
Two important aspects of any forecast:
1. Expected level of demand – may depend on trend or seasonal
variation.
2. Accuracy – the degree to which the forecast might deviate from the
actual value (forecast error).
Two uses of forecasts
Plan the system → long-range decisions (capacity, facilities, location).
Plan the use of the system → short/intermediate-range decisions
(inventory, workforce, scheduling, budgeting).
Features Common to All Forecasts
1. Techniques assume the same underlying causal system that existed
in the past will continue into the future.
2. Forecasts are not perfect — random variation always causes some
error.
3. Forecasts for groups of items are more accurate than for individual
items (errors cancel out).
4. Accuracy decreases as the forecast time horizon increases.
2. ELEMENTS OF A GOOD FORECAST
A good forecast should be:
1. Timely — must cover the lead time needed to react (e.g., toy retailer
needs a forecast made early enough to cover the 2-month supplier
lead time before Christmas).
2. Accurate — and the degree of accuracy stated.
3. Reliable — should work consistently.
4. In meaningful units (dollars, units, machine-hours, etc.).
5. In writing.
6. Simple to understand and use.
7. Cost-effective (benefits > costs).
3. STEPS IN THE FORECASTING PROCESS
1. Determine the purpose of the forecast.
2. Establish a time horizon.
3. Obtain, clean, and analyze appropriate data (remove outliers/incorrect
data).
4. Select a forecasting technique.
5. Make the forecast.
6. Monitor the forecast (revise if it is not performing well).
4. FORECAST ACCURACY
Forecast Error = Actual − Forecast
$$e_t = A_t - F_t$$
Positive error → forecast was too low.
Negative error → forecast was too high.
Small error → high accuracy. Large error → low accuracy.
Why accuracy matters
Accurate forecasts help organizations:
Make better production plans
Maintain the right inventory levels
Reduce costs from overproduction/shortages
Improve customer satisfaction
5. SUMMARIZING FORECAST ACCURACY — MAD,
MSE, MAPE
Measure Formula What it does
MAD (Mean Average size of errors
$\dfrac{\sum \lvert A_t - F_t
Absolute (all errors weighted
\rvert}{n}$
Deviation) equally)
Squares errors →
MSE (Mean $\dfrac{\sum (A_t - F_t)^2}{n-
penalizes large errors
Squared Error) 1}$
heavily
$\dfrac{\sum Expresses error as a
MAPE (Mean
\left[\dfrac{\lvert A_t- % of actual — good
Absolute
F_t\rvert} for comparing across
Percent Error)
{A_t}\times100\right]}{n}$ different scales
Quick guide:
MAD → easiest to compute, treats every error the same.
MSE → best when large errors are especially costly (e.g., stockouts).
MAPE → best when you need to compare accuracy across
items/products of very different sizes.
🔢 WORKED EXAMPLE 1 — Compute MAD, MSE, MAPE
Data: Actual & Forecasted number of accounts serviced (8 periods)
Error |Error|/Actual
Period Actual Forecast |Error| Error²
(A−F) ×100
1 217 215 2 2 4 0.92%
2 213 216 −3 3 9 1.41%
3 216 215 1 1 1 0.46%
4 210 214 −4 4 16 1.90%
5 213 211 2 2 4 0.94%
6 219 214 5 5 25 2.28%
7 216 217 −1 1 1 0.46%
8 212 216 −4 4 16 1.89%
Σ −2 22 76 10.26%
Calculations (n = 8):
$$MAD = \frac{22}{8} = 2.75 \text{ units}$$
$$MSE = \frac{76}{8-1} = \frac{76}{7} = 10.86 \text{ units}^2$$
$$MAPE = \frac{10.26%}{8} = 1.28%$$
Interpretation:
On average, the forecast is off by about 2.75 accounts (MAD).
MSE = 10.86 shows that large errors (like the error of 5 in period 6)
are penalized more — the value is much bigger than MAD because it's
squared.
MAPE = 1.28% means the forecast is, on average, wrong by only
1.28% of the actual value — this is a very accurate forecast in relative
terms.
🔢 WHY THREE DIFFERENT MEASURES? — Illustrative
Comparisons
(a) MAD vs MSE — Does MAD hide big mistakes?
Suppose forecast errors over 5 periods are: 1, 1, 1, 1, 16
$$MAD = \frac{1+1+1+1+16}{5} = \frac{20}{5} = 4$$
$$MSE = \frac{1^2+1^2+1^2+1^2+16^2}{5} = \frac{1+1+1+1+256}{5} =
\frac{260}{5} = 52$$
Interpretation: MAD (=4) makes the errors look small and acceptable on
average. But MSE (=52) is huge because squaring 16 gives 256 — MSE
"shouts" that there was one serious mistake that MAD hides. This is
critical in situations like a supermarket forecast, where one big miss (a
stockout day) can mean major lost sales — MSE flags that danger; MAD
doesn't.
(b) MSE vs MAPE — Which error is "worse" relatively?
Product Actual Forecast Error
A 100 90 10
B 20 10 10
Squared error for both A and B = 10² = 100 → MSE says both forecasts
are equally bad.
But using MAPE:
$$\text{Product A: } \frac{10}{100}\times100 = 10%$$
$$\text{Product B: } \frac{10}{20}\times100 = 50%$$
Interpretation: Although the absolute error is the same (10 units) for both
products, being off by 10 units when actual demand is only 20 is a much
bigger relative mistake (50%) than being off by 10 when actual demand is
100 (10%). MAPE captures this — MSE cannot. This is why MAPE is
preferred when comparing accuracy across products/items of very
different scale.
6. APPROACHES TO FORECASTING
Two general approaches:
Approach Description
Subjective inputs (opinions, judgment, experience). Used
Qualitative when historical data is unavailable or conditions are
changing (new products, political/economic shifts).
Analyzing objective/historical data — either time-series
Quantitative (project past data forward) or associative/causal models
(use other variables to predict demand).
Qualitative Techniques
Executive opinions – group of senior managers develop a forecast
together. Risk: one person's opinion may dominate.
Salesforce opinions – sales staff are close to customers but may be
overly optimistic/pessimistic based on recent experience.
Consumer surveys – direct from the customer, but expensive &
subject to irrational responses.
Delphi method – iterative, anonymous questionnaire process among
experts to reach consensus (used for long-range/technological
forecasts where no hard data exists).
🔢 WORKED EXAMPLE — Delphi Method
Question: When will 5G reach ≥50% household usage in Kaliganj,
Bangladesh?
Round 1 (independent, initial estimates):
| Expert | A | B | C | D | E |
|---|---|---|---|---|---|
| Estimate | 2030 | 2032 | 2035 | 2029 | 2034 |
Round 1 Range: 2029 – 2035
Round 2 (after seeing group summary, experts revise):
| Expert | A | B | C | D | E |
|---|---|---|---|---|---|
| Revised | 2031 | 2032 | 2034 | 2030 | 2033 |
Round 2 Range: 2030 – 2034
Interpretation: The range narrowed from 7 years to 4 years after just one
round of feedback — this is the whole point of Delphi: anonymous,
iterative feedback moves independent experts toward a converged
consensus without letting one dominant personality control the outcome.
Final conclusion: 5G likely reaches 50% household penetration in Kaliganj
between 2030 and 2034.
7. TIME-SERIES FORECASTING
A time series = a time-ordered sequence of observations (hourly, daily,
weekly, etc.)
Time-series behaviors to look for (plot the data first!):
1. Trend – long-term upward/downward movement.
2. Seasonality – short-term regular variation tied to the calendar/time of
day.
3. Cycles – wave-like variations lasting more than one year.
4. Irregular variation – due to unusual, one-off events (should be
removed from data).
5. Random variation – residual variation left after everything else is
accounted for.
Naive Method
The forecast for the next period = last period's actual value (or, with trend,
last value + recent change).
Advantages: virtually no cost, easy, quick, easy to understand.
Disadvantage: not very accurate — but it's a useful baseline to judge
whether a more sophisticated (and costly) method is actually worth
using.
8. AVERAGING TECHNIQUES
Averaging techniques smooth out random variation ("white noise") in the
data. Three methods:
(A) Simple Moving Average
$$F_t = MA_n = \frac{A_{t-1}+A_{t-2}+\dots+A_{t-n}}{n}$$
Every one of the last n periods gets equal weight.
Fewer periods in the average → more responsive (reacts quickly, but
noisier).
More periods → smoother, but slower to react to real changes.
🔢 Worked Example — 3-period Moving Average
Period Demand
1 42
2 40
3 43
4 40
5 41
$$F_6 = \frac{43+40+41}{3} = \frac{124}{3} = 41.33$$
If actual demand in period 6 = 38:
$$F_7 = \frac{40+41+38}{3} = \frac{119}{3} = 39.67$$
Interpretation: Each time a new actual value comes in, the average
"moves" — the oldest value drops off and the newest is added. Notice the
moving average forecast lags the actual data (it's smoother but slower to
catch turning points).
(B) Weighted Moving Average
$$F_t = w_t(A_t) + w_{t-1}(A_{t-1}) + \dots + w_{t-n}(A_{t-n})$$
Weights must sum to 1.00, with the heaviest weight on the most
recent period (recent data is usually more informative — like
yesterday's weather predicting tomorrow better than 3-week-old
weather).
🔢 Worked Example — Weighted Average (weights: .40, .30, .20, .10)
Period Demand
1 42
2 40
3 43
4 40
5 41
(a) Forecast for period 6 (weight .40→most recent=period5, .30→period4,
.20→period3, .10→period2):
$$F_6 = .10(40) + .20(43) + .30(40) + .40(41)$$
$$F_6 = 4 + 8.6 + 12 + 16.4 = 41.0$$
(b) If actual demand in period 6 = 39, forecast period 7:
$$F_7 = .10(43) + .20(40) + .30(41) + .40(39)$$
$$F_7 = 4.3 + 8 + 12.3 + 15.6 = 40.2$$
Interpretation: Compared to the simple moving average (41.33 for the
same data), the weighted average shifts slightly because it gives more
importance to the most recent period. The advantage over a simple
moving average is that it reacts faster to recent changes; the
disadvantage is that the choice of weights is arbitrary (usually found via
trial and error).
(C) Exponential Smoothing
The most widely used averaging technique — a sophisticated weighted
average that only needs two numbers: the previous forecast and the
latest actual value.
$$F_t = F_{t-1} + \alpha(A_{t-1} - F_{t-1})$$
equivalently:
$$F_t = (1-\alpha)F_{t-1} + \alpha A_{t-1}$$
Where:
$\alpha$ (alpha) = smoothing constant, 0 to 1
$(A_{t-1}-F_{t-1})$ = the previous forecast error
Choosing α:
| Situation | α value | Why |
|---|---|---|
| Random/unusual spike (e.g., one hot day) | Low α (e.g., 0.05–0.10) | We
don't trust the spike — assume no real trend, just noise → small
adjustment |
| Real, persistent trend (e.g., temperature rising every day) | High α (e.g.,
0.40+) | We believe the change is real → forecast should move quickly to
catch up |
Commonly used range: α = 0.05 to 0.50. Low α = stable series, more
smoothing. High α = more responsive, less smoothing.
🔢 Worked Example — Exponential Smoothing Comparison
Compare Naive, 2-period Moving Average, and Exponential Smoothing (α
= 0.10) for periods 3 through 11:
| Demand data (Period 1–11): | 42, 40, 43, 40, 41, 39, 46, 44, 45, 38, 40 |
Exp.
2-
Period, Naive Naive MA Smooth Exp
Demand period
t F Error Error F Error
MA F
(α=.10)
1 42 — — — — — —
2 40 42 −2 — — 42 −2
3 43 40 3 41 2 41.80 1.20
4 40 43 −3 41.5 −1.5 41.92 −1.92
5 41 40 1 41.5 −0.5 41.73 −0.73
6 39 41 −2 40.5 −1.5 41.66 −2.66
7 46 39 7 40.0 6.0 41.39 4.61
8 44 46 −2 42.5 1.5 41.85 2.15
9 45 44 1 45.0 0.0 42.07 2.93
10 38 45 −7 44.5 −6.5 42.36 −4.36
11 40 38 2 41.5 −1.5 41.92 −1.92
Sample calculation (Exponential Smoothing):
Starting forecast for period 2 = naive value = 42 (this is the standard
starting-point convention).
$$F_3 = 42 + 0.10(40-42) = 42 - 0.20 = 41.80$$
$$F_4 = 41.80 + 0.10(43-41.80) = 41.80+0.12 = 41.92$$
...and so on, each new forecast = old forecast + 10% of the last error.
Error summary (using periods 3–11, n = 9 for MAD/MAPE, n−1=8 for
MSE):
Measure Naive 2-period MA Exp. Smoothing (α=.10)
MAD 3.11 2.33 2.50
MSE 16.25 11.44 8.73
MAPE 7.49% 5.64% 5.98%
Result interpretation:
If MAD is the criterion → 2-period moving average wins (lowest
average absolute error).
If MSE is the criterion → Exponential smoothing wins (best at
controlling large errors).
If MAPE is the criterion → 2-period moving average wins again
(lowest % error).
The naive method performs worst on all three measures here — but
it's also the cheapest/simplest, so a manager must weigh the small
accuracy gain of the other methods against their added complexity
and cost.
Note: With different smoothing constants or a different-length moving
average, the ranking could change — there is no single "best" method
for all situations.
9. ASSOCIATIVE FORECASTING TECHNIQUES —
SIMPLE LINEAR REGRESSION
Associative (causal) techniques use a predictor variable (x) to forecast a
dependent variable (y). The most common method is simple linear
regression — fitting a straight line that minimizes the sum of squared
vertical deviations of the data points from the line (the least squares
criterion).
$$y_c = a + bx$$
where:
$y_c$ = predicted (dependent) variable
$x$ = predictor (independent) variable
$b$ = slope of the line
$a$ = value of $y_c$ when $x = 0$ (the y-intercept)
Formulas for the coefficients:
$$b = \frac{n\sum xy - \sum x \sum y}{n\sum x^2 - (\sum x)^2}$$
$$a = \frac{\sum y - b\sum x}{n} \quad \left(= \bar{y} - b\bar{x}\right)$$
Standard Error of Estimate (measures how scattered the data points are
around the regression line — the smaller it is, the more accurate the
predictions):
$$S_e = \sqrt{\frac{\sum(y-y_c)^2}{n-2}}$$
Correlation coefficient (r) measures the strength/direction of the linear
relationship (ranges from −1.00 to +1.00). r² ("coefficient of
determination") tells you what % of the variation in y is explained by x.
Rule of thumb: r² ≥ 0.80 → good predictor; r² between 0.25–0.80 →
moderate predictor; r² ≤ 0.25 → poor predictor.
🔢 WORKED EXAMPLE — "Healthy Hamburgers" Sales vs. Profit
A chain of 12 stores. x = Unit Sales ($ millions), y = Profits ($ millions).
Goal: predict profit for a store with sales of $10 million.
Store x (Sales) y (Profit) xy x²
1 7 0.15 1.05 49
2 2 0.10 0.20 4
3 6 0.13 0.78 36
4 4 0.15 0.60 16
5 14 0.25 3.50 196
Store x (Sales) y (Profit) xy x²
6 15 0.27 4.05 225
7 16 0.24 3.84 256
8 12 0.20 2.40 144
9 14 0.27 3.78 196
10 20 0.44 8.80 400
11 15 0.34 5.10 225
12 7 0.17 1.19 49
Σ 132 2.71 35.29 1796
n = 12, $\bar{x}=11$, $\bar{y}=0.2258$
Step 1 — Slope (b):
$$b = \frac{12(35.29) - (132)(2.71)}{12(1796) - (132)^2} = \frac{423.48 -
357.72}{21552-17424} = \frac{65.76}{4128} = 0.0159$$
Step 2 — Intercept (a):
$$a = \frac{2.71 - 0.0159(132)}{12} = \frac{2.71-2.0988}{12} =
\frac{0.6112}{12} = 0.0506$$
Regression equation:
$$y_c = 0.0506 + 0.0159x$$
Step 3 — Predict profit for sales of $10 million:
$$y_c = 0.0506 + 0.0159(10) = 0.0506+0.159 = 0.2099 \text{ (≈
$209,900)}$$
Step 4 — Forecast & error for every store (ŷ = a + bx, error = y − ŷ):
x y (actual) ŷ (forecast) Error (y − ŷ)
7 0.15 0.1621 −0.0121
x y (actual) ŷ (forecast) Error (y − ŷ)
2 0.10 0.0825 0.0175
6 0.13 0.1462 −0.0162
4 0.15 0.1143 0.0357
14 0.25 0.2736 −0.0236
15 0.27 0.2896 −0.0196
16 0.24 0.3055 −0.0655
12 0.20 0.2418 −0.0418
14 0.27 0.2736 −0.0036
20 0.44 0.3692 0.0708
15 0.34 0.2896 0.0504
7 0.17 0.1621 0.0079
Step 5 — Standard Error of Estimate:
Sum of squared errors $\sum(y-y_c)^2 \approx 0.01659$
$$S_e = \sqrt{\frac{0.01659}{12-2}} = \sqrt{0.001659} \approx 0.0407
\text{ ($ million)}$$
Step 6 — Correlation & r²:
$$r \approx 0.917 \qquad r^2 \approx 0.840$$
Interpretation:
The slope b = 0.0159 means: for every additional $1 million in sales,
profit is predicted to rise by about $15,900.
The intercept a = 0.0506 is just where the line crosses the y-axis (x=0)
— it should not be interpreted as "profit with zero sales"; it's only
meaningful for establishing the height of the line, since x=0 is outside
the range of the observed data (sales ranged from 2 to 20).
Predicted profit at $10 million sales = $209,900 — this is only reliable
within the observed sales range (2–20); extrapolating far outside this
range is risky.
r² = 0.840 means 84% of the variation in profit is explained by sales
volume — this is a strong predictor (≥0.80 rule of thumb), so linear
regression is a good, credible forecasting tool here.
Standard error of estimate (Se ≈ 0.0407 million) tells us how tightly
the actual profits cluster around the regression line — a small Se
relative to the y-values means the regression line gives fairly precise
predictions; a large Se would mean predictions could be off by a wide
margin.
Conditions for a valid regression-based forecast:
1. Variation around the line should be random (no visible trend/cycle
pattern in the residuals).
2. Deviations around the line should be normally distributed.
3. Predictions should only be made within the range of the observed x
values.
10. QUICK-REFERENCE FORMULA SHEET
Concept Formula
Forecast Error $e_t = A_t - F_t$
MAD $\dfrac{\sum \lvert A_t-F_t\rvert}{n}$
MSE $\dfrac{\sum (A_t-F_t)^2}{n-1}$
$\dfrac{\sum\left[\frac{\lvert A_t-F_t\rvert}
MAPE
{A_t}\times100\right]}{n}$
Naive forecast
$F_t = A_{t-1}$
(stable)
Naive forecast
$F_t = A_{t-1} + (A_{t-1}-A_{t-2})$
(with trend)
Concept Formula
Moving Average $F_t = \dfrac{A_{t-1}+A_{t-2}+\dots+A_{t-n}}{n}$
Weighted Moving $F_t = w_t A_t + w_{t-1}A_{t-1}+\dots+w_{t-n}A_{t-
Average n}$ (weights sum to 1)
Exponential
$F_t = F_{t-1}+\alpha(A_{t-1}-F_{t-1})$
Smoothing
Linear Regression $b = \dfrac{n\sum xy - \sum x\sum y}{n\sum x^2-
slope (\sum x)^2}$
Linear Regression
$a = \dfrac{\sum y - b\sum x}{n}$
intercept
Regression
$y_c = a+bx$
forecast
Standard Error of
$S_e=\sqrt{\dfrac{\sum(y-y_c)^2}{n-2}}$
Estimate
11. KEY TAKEAWAYS (for exam answers on
"interpretation")
Always plot the data first — this tells you whether you're dealing with a
stable series, trend, or seasonality, and therefore which technique is
appropriate.
A lower MAD/MSE/MAPE = a more accurate forecast, but the "best"
measure to use depends on the situation:
Use MSE when large errors are very costly (e.g., stockouts,
safety).
Use MAPE when you need to compare forecast accuracy across
items of different scale/size.
Use MAD when you want a simple, easy-to-communicate
measure.
Moving average vs weighted average vs exponential smoothing: all
three are averaging techniques that smooth random variation; the
difference is how much weight recent data gets. Exponential
smoothing is preferred in practice because it only requires 2 numbers
(last forecast + last actual) — no need to store a long history of data
like a moving average does.
The smoothing constant (α) controls the trade-off between stability
(low α, trusts history more) and responsiveness (high α, trusts the
newest data point more, reacts fast to real change but also to noise).
Random variation can never be eliminated — only distinguished from
real variation as best as possible. No forecasting technique produces
a perfect forecast.
Regression is different from time-series methods (naive, moving
average, exponential smoothing): time-series methods only look at
the variable's own past values, while regression uses a separate
predictor variable (like sales, advertising spend, or unemployment
rate) to explain and predict the variable of interest. Always check r² to
judge whether the predictor variable is actually a good one before
trusting the regression forecast.