Homework 7 — One strategy, every pitfall answered with evidence

Due: Tuesday 17 November 2026, 23:59 · folder week-07/hw/ in your repo

Goal

Backtest one trading rule honestly on at least two assets, with costs and a walk-forward, and — the real deliverable — answer the pitfall checklist with evidence from your own notebook. Then add one text-derived feature and show whether it earns its place. Everything is an educational simulation: no live trading, no advice.

Pick a rule (one)

• A momentum or mean-reversion rule with one or two parameters (as in Block B), on your capstone asset plus one other.
• A volatility-targeting rule: position size = target vol / realised vol, long only. Message me first if you choose this; the cost model changes.
• The week-3 classifier turned into a rule: long when P(up) > threshold, flat otherwise, using your HW3 model.
• Your own, if it is a position decided at close t and held from t+1. Message me first.

Deliverables (in week-07/hw/)

1. hw7.ipynb — runs from a clean runtime. Must contain, in this order:
   • data acquisition with the three-tier loader; at least two assets, from 2018 or the earliest available; snapshot saved;
   • the rule as a function, and the backtest pipeline with the shift(1) commented in the cell where it happens;
   • a statistics table (annual return, vol, Sharpe, max drawdown, turnover per year) at 0, 10 and 25 bps per unit traded, per asset;
   • a walk-forward parameter choice: fit on an expanding or rolling window, apply to the next block, report the out-of-sample Sharpe next to the best in-sample Sharpe and the number of parameter values tried;
   • a by-period table (by year, and at least one split you name as a regime);
   • one text-derived feature — Block D’s headline scorer on your own headline set, or any dated text you can obtain (press releases, protocol announcements, exchange status posts, your own notes) — with its published column, added to the rule or to the classifier, shown against the baseline without it, net of the same costs;
   • the pitfall checklist as a markdown table (below), each row answered with the cell number that is the evidence.
2. memo.md (400–600 words) with these headings:
   • The rule and the assets
   • Honest result (one table: per asset, Sharpe at 10 bps, walk-forward OOS Sharpe, worst year)
   • Pitfall checklist — one line per row: look-ahead, survivorship, multiple testing, costs, regime, data snooping; each answered “not affected because…” or “affected, and here is the size of it…”
   • The text feature — what it adds beyond the price, or that it adds nothing, with the number
   • What would have to be true for anyone to act on this, and the monitoring metric and kill-switch threshold you would set before go-live
   • AI-use disclosure — including every parameter value the assistant tried on your behalf

Pitfall checklist (copy into the notebook)

pitfall	question	evidence (cell)
look-ahead	is every position decided with data ≤ the previous close?
survivorship	is the asset list the list you would have had on the start date?
multiple testing	how many parameter values were tried in total, by you and by the assistant?
costs	is the cost charged per unit traded, and is the level stated?
regime	does the result hold in more than one period?
data snooping	did you look at the out-of-sample period before choosing anything?

Constraints

• No parameter chosen on the full sample. If the assistant proposes a grid search without a walk-forward, that is your “what the assistant got wrong” this week — say so in the memo.
• Costs per unit traded, never per day held. A rule with turnover above 100 per year must be shown at 25 bps too.
• The text feature must be dated by availability, and the memo must state where the dates come from. If you cannot obtain availability dates, say so and use Block D’s synthetic set, labelled.
• Two assets minimum. A rule that works on one and not the other is a valid and common result — report it as such.

Scope

4–6 hours. The strategy is the excuse; the checklist with evidence and the memo are what I read first. A memo that says “this rule has no edge after costs and here are the six pieces of evidence” is a good memo.

Self-check

• Restart and run all: no errors; first data line prints [live] or [snapshot].
• shift(1) is present and commented.
• Statistics at 0/10/25 bps, per asset.
• Walk-forward OOS Sharpe next to best in-sample, with N tried.
• By-year table present.
• Text feature vs baseline, same costs, availability dates stated.
• Checklist table filled with cell numbers; memo has all six headings.
• Monitoring metric and kill-switch threshold stated; AI-use disclosure lists parameters tried.

How this feeds the capstone

This is the strategy or decision-rule layer of your pipeline, and the checklist is the first half of the risk and controls canvas (week 9 adds the responsible-AI half). Week 8 supplies on-chain features you can add to the same backtest. If your capstone asset is a price series, this notebook becomes a section of the portfolio notebook as it stands.