Internship Screening Assignment —
Trading RL System Design
Format: Written responses only. No coding required.
Submission: Single PDF or Markdown document, 4–8 pages
Domain: Algorithmic trading using Machine learning + Reinforcement Learning
---
## Purpose
This is a design-thinking screening test, not a coding test.
We want to know: can you reason clearly about an ML system in a complex domain (trading) under
realistic constraints? Can you make defensible design choices? Can you spot weaknesses in your own
thinking?
There are no perfect answers. We are evaluating how you think, not what you conclude.
---
## Instructions
1. Read the scenario below.
2. Answer all 5 questions in writing.
3. Don't need to write code. Pseudocode is allowed where it clarifies a point, but most answers should be
prose.
4. Use diagrams (hand-drawn photos, [Link] exports, ASCII) wherever they help.
5. Submit a single document.
You may use any reference (Google, books, AI assistants). But every answer must reflect your own
reasoning. We will ask follow-up questions in the next round, so make sure you can defend everything
you write.
---
## The Scenario
You are asked to design a system that:
1. Learns price-action patterns from historical OHLCV data of an asset using Machine learning. The
goal is to discover recurring market structures (regimes, patterns, level types) without being told what to
look for.
2. Trains a reinforcement learning agent that uses these learned patterns as part of its observation space.
The agent decides when to enter trades, where to place stop-losses, and where to place take-profits.
3. Targets high risk-reward trades: tight stop-losses (small losses), ambitious take-profits (large wins).
The agent should learn to wait for high-quality setups rather than trade frequently.
You are the lead engineer. You must hand a design to a junior who will implement it.
---
## Questions
### Question 1 — Problem Decomposition
You have 6 months to ship this. Break the project into 3–5 milestones, each with:
- A clear deliverable
- A success criterion (how you know the milestone is complete)
- The biggest risk in that milestone
Format your answer as a simple table or list.
---
### Question 2 — Pattern Learning Design
The first phase uses unsupervised learning to discover patterns in price data.
Answer:
(a) What specifically would you cluster, embed, or model? Be concrete — not "I would use deep learning
on price data." Specify: input features, model type, what the output represents.
(b) Which learning approach would you use for this phase — supervised, unsupervised, self-supervised, or
rule-based feature engineering? Justify your choice. Specifically address: why this approach is right, and
why you rejected the others.
(c) How would you validate that the unsupervised model learned something useful, BEFORE plugging it
into the RL agent? You need a sanity check.
---
### Question 3 — RL Action Space and Reward Design
The agent must decide:
- Whether to enter a trade (Hold / Long / Short/ Close)
- Where to place the stop-loss
- Where to place the take-profit
Answer:
(a) Design the action space. Be specific about types (discrete, continuous), ranges, and what each action
means. Why these choices?
(b) Design the reward function. The user explicitly wants tight SL and ambitious TP for high RR. Your
reward function must encourage this. Give the formula and explain each term.
(c) Identify ONE way the agent could hack your reward function to maximize reward without actually
trading well. How would you prevent this?
---
### Question 4 — Evaluation and Deployment
Answer:
(a) How would you evaluate the agent BEFORE deploying it live? What are the metrics, and what is your
acceptance threshold?
(b) You're about to deploy with $50 of real money. List 3 specific things you would do (or build, or
check) the day before going live that you would NOT have bothered doing during development.
---
### Question 5 — Self-Critique (15 minutes, ~1 page)
Look at your answers to Questions 1–4. Now critique them.
(a) Identify the 3 weakest decisions in your design. For each, explain:
- Why it's weak
- What would have to be true for your weak decision to actually be the right choice
(b) What is one question you wish was on this assessment that wasn't? What would your answer be?
---
## Evaluation Criteria
Design Thinking
- Strong answer: Specific choices with stated tradeoffs. Acknowledges alternatives.
- Weak answer: Generic suggestions. "Use neural network." No tradeoffs mentioned.
- Implementation Quality
- Strong answer: Pseudocode/structure shows the candidate could actually build this.
- Weak answer: Hand-waves over hard parts. "Then we train the model."
- End-to-End Completeness
- Strong answer: Every question answered. Connections between answers are coherent.
- Weak answer: Skips parts. Q3 design contradicts Q2 design.
- Evaluation Rigor
- Strong answer: Specific metrics. Realistic thresholds. Acknowledges what they don't measure.
- Weak answer: "Sharpe ratio should be high." No threshold given.
- Self-Awareness
- Strong answer: Q5 critique identifies real, specific flaws.
- Weak answer: Q5 says "I would tune hyperparameters more."
---
## Submission Process
Submit your assignment on Internshala
We respond within 3 business days. If you pass, we schedule a 30-minute follow-up call to discuss your
answers.
---
## A Note on AI Tools
You can use Claude / ChatGPT / Copilot. We use them ourselves.
But: every answer must be defendable in conversation. In the follow-up call, we will ask "why did you
say X in Question 2?" If your answer is genuinely yours, you can defend it.
Strong candidates use AI to validate their thinking, not to replace it.
---
Good luck. We genuinely want you to do well — finding strong candidates is hard, and we're rooting for
you.