Alternative Data Trading Course: Turn Attention and Filings Into Tests

Alternative data is easy to collect and unusually easy to misuse. A chart of search interest beside a stock price can look persuasive even when the dates are misaligned, the ticker mapping is wrong, and the interesting period was selected after the outcome was known. A serious alternative data trading course should therefore spend less time celebrating exotic feeds and more time teaching how a noisy observation becomes a testable claim.

This article is a curriculum blueprint. It differs from our existing guide to alternative data for event-driven research, which surveys useful data categories and applications. Here the question is narrower and more practical: what should a student learn, in what order, before calling an alternative dataset a signal?

Start with a mechanism, not a feed

A dataset is not an edge merely because most traders do not have it. Wikipedia views, app rankings, web traffic, filing frequency, and satellite images are measurements. The research begins when you state who is acting, why their behavior reaches the measurement, and why that behavior could precede a repricing. Without that chain, a correlation is only a prompt for another test.

A useful opening assignment starts with a plain sentence. For example, unusual attention may precede higher volatility because an unresolved company event is drawing a wider audience before formal news arrives. That claim predicts magnitude, not direction. It also admits several ways to be wrong: attention may follow the price move, bots may distort the series, or the effect may disappear after controlling for earnings and sector.

Teach the point-in-time audit before modeling

Alternative datasets often change after first publication. A company can change its name, a web provider can revise history, and a security can disappear after delisting. The student must learn to distinguish event time, collection time, vendor publication time, and the time a value became available to the strategy. If those clocks are collapsed into one date column, a model can quietly receive tomorrow's information.

The course should make students inspect raw rows before writing features. They should find missing days, duplicated entities, ticker changes, stale values, and revisions. They should also build a point-in-time universe rather than applying today's surviving tickers to old dates. This work is not glamorous, but it is where research credibility is won.

Turn raw observations into comparable features

Raw counts rarely compare cleanly across companies. A thousand page views may be extraordinary for a small issuer and routine for a household name. Students should normalize each series against its own recent history, then ask whether the result survives a cross-sectional comparison. Rolling medians, robust z-scores, percentage changes, and percentile ranks are useful, provided every lookback uses only information available at the time.

Conditioning comes next. Attention can rise around earnings, financing announcements, index changes, or broad sector news. Filing activity naturally differs by company size and industry. A course exercise should split the sample by scheduled catalysts, market-cap bands, sectors, and market regimes. The goal is not to maximize a backtest. It is to learn whether the proposed mechanism remains visible after obvious explanations are removed.

Students should also learn that an accurate prediction can be economically useless. A feature may identify volatile names only after spreads have widened and prices have moved. A vendor may publish a daily observation after the intended decision time. Turnover may consume a small statistical advantage. Each exercise should therefore end with a latency and implementation audit. State when the signal becomes observable, when a trade could reasonably occur, what universe remains liquid, and how much of the measured effect is left after delay. This turns an interesting relationship into an honest decision about whether further research is justified.

• Timestamp test. Rebuild every feature as it would have existed before the trade decision.
• Baseline test. Compare the alternative feature with price, volume, volatility, and scheduled events.
• Stability test. Recompute results by year, sector, and signal-strength bucket.
• Friction test. Include delay, turnover, liquidity, and any vendor availability constraint.

Make students attack the attractive result

The most dangerous assignment is a leaderboard where the best Sharpe ratio wins. It rewards repeated searching and hides the number of failed specifications. A better assignment fixes the hypothesis and evaluation plan first. Students report every sensible variant, including the ones that weaken the thesis, and explain which result was expected before the data was examined.

Reproducibility should be assessed directly. Another student receives the frozen hypothesis, data dictionary, and repository, then attempts to rebuild the principal table without verbal help. Any ambiguity becomes part of the grade: undocumented exclusions, local files, changing identifiers, or unexplained feature choices. This peer replication is valuable because alternative data projects accumulate quiet judgment calls. Making those choices visible protects the result and teaches students to write research that survives beyond their own notebook.

They should also separate prediction from economic use. An attention spike may forecast absolute movement yet provide no directional edge. That can still matter for watchlist construction or volatility research. Forcing every dataset into a long or short call discards honest findings and encourages invented narratives.

The course outcome should be a defended research object

A finished project should include a mechanism statement, a point-in-time data audit, reproducible feature code, conditioned result tables, and a written failure analysis. The Attention as Data lesson provides a concrete starting point using Wikipedia views, while the quant trading curriculum hub places it inside the wider research sequence.

That outcome is more valuable than a catalog of unusual feeds. Vendors change and fashionable datasets rotate. The durable skill is learning to turn a messy measurement into a claim that can be dated, tested, conditioned, and rejected. That is what an alternative data trading course should teach.