NLP and Alternative Data

Text is the one domain where modern deep learning has unambiguously changed quant practice, and "alternative data" is where new edges are hunted now that price-based factors are crowded. This closing lesson covers both — methods first, then the pitfalls that are special to this data.

Text into numbers: the ladder

• Dictionary methods: count positive/negative words. Crude but honorable — the finance-specific Loughran–McDonald lexicon exists because general sentiment dictionaries misfire on filings ("liability," "tax" aren't negative sentiment in 10-Ks); knowing that citation signals domain literacy.
• Bag-of-words + linear models: TF-IDF features into the logistic regression of the classification lesson; interpretable, strong baselines, still competitive on many filing-classification tasks.
• Embeddings and transformers: words (then sentences, then documents) as dense vectors capturing meaning; fine-tuned transformers read earnings calls and news with near-human comprehension at machine scale. The modern pipeline: embed each document, then feed embeddings to the boring, regularized models this course keeps recommending — the deep part does reading, the simple part does predicting.

What the research literature consistently finds: text signals are real but fast-decaying (news is priced in hours-to-days) and capacity-constrained — the feature-engineering lesson's turnover-vs-cost analysis decides viability more often than model quality does.

Alternative data: the menu and the economics

Card-transaction panels (nowcast retailer revenue), satellite imagery (parking lots, oil storage shadows), app usage, job postings, shipping traffic. The evaluation framework matters more than the catalog — for any dataset ask: coverage (how many tradeable names?), history (enough to validate — many vendors offer 3 noisy years), timeliness vs the market (who else has it, and faster?), cost vs capacity (a $500k dataset needs real capacity to pay), and mapping (joining "merchant" to "ticker" is unglamorous and is where half the projects die).

Pitfalls specific to this data

Beyond the standard seven backtesting sins, alt data adds its own:

• Backfill and survivorship at the vendor: the dataset's "history" was often reconstructed after the vendor cleaned house — panels that changed composition, tickers mapped with today's knowledge. Point-in-time discipline applies to the vendor's process, not just your own.
• Panel drift: a card panel's demographics shift as the app's user base shifts; the signal you validated isn't the signal you're now receiving. Monitor the data, not only the model (the ML-in-markets lesson's monitoring, extended upstream).
• Crowding: the moment a dataset is productized, its alpha decays toward its cost (the adaptive-markets logic) — first-buyer advantage is a real but wasting asset. The equilibrium is uncomfortable and worth stating: widely-sold alt data converges to expensive beta.
• LLM-era leakage: a model pretrained on the internet has read the future relative to any historical backtest date — evaluating an LLM-derived signal on 2019 news when the model trained on 2024 text is lookahead bias in its newest costume. The field's current best practice (time-partitioned models, strict cutoffs) is imperfect and honest answers say so.

The interview version

"You're offered a credit-card panel claiming to nowcast retail revenues. How do you evaluate it?" — Mapping and coverage audit; point-in-time reconstruction of the vendor's history; correlation with subsequent revenue surprises out-of-sample; decay and crowding check (how many clients?); cost against realistic capacity; then a small live paper-trade before capital. "Would an LLM reading news beat a word-count model?" — At reading, yes; at predicting, only if the extra comprehension survives costs, decay, and the leakage audit — and the burden of proof is on the expensive model. Same epistemology as the whole course, aimed at the shiniest data.