Add boxoffice-signal-ml-starter: predicting outcomes from public conversation with Cortex AISQL

[snowflake-mike-MKTG-IND]

Summary

Adds samples/boxoffice-signal-ml-starter — a method kit for predicting a commercial outcome from public pre-release conversation, guided end to end by Cortex Code. The worked example is opening-weekend box office, but the pattern applies anywhere public conversation precedes a commercial outcome (game launches, product releases, ticketed events).

The interesting part is not the domain, it's the discipline around the AI:

• Cortex AISQL intent scoring, not sentiment counting. One consolidated TRY_COMPLETE call classifies each comment into four viewing-intent labels. Independent per-label yes/no calls over-fire — that lesson is baked into the prompt and documented.
• An evaluation harness for the classifier itself (sql/30_intent_eval.sql): macro-F1 plus a per-class confusion matrix against hand labels. On the reference build, the production scorer measured macro-F1 0.65 while a revised prompt reached 0.85 on the same labels — a gap that is invisible from reading output.
• As-of feature engineering. Demand percentiles at −21/−14/−7/−3 day horizons where no feature can see past its own prediction date, plus three tripwires in sql/10 that catch it when the filter silently breaks.
• Strict walk-forward temporal validation with a distributional regressor (CatBoost + ElasticNet blend, residual-mixture prediction intervals, calibrated demand-forward flag), and pedigree features gated behind demand so the model can't lean on hype the audience isn't backing.

Runs in ten minutes with no data

cd samples/boxoffice-signal-ml-starter
pip install -r model/requirements.txt
python model/train_ow_model.py --sample

Synthetic data, no Snowflake connection required — a reviewer can verify the modeling path end to end before wiring up any source.

Data, terms, and privacy

No data is included and no provider is named or endorsed. The sources/ dossiers describe each signal by its characteristics — behavior, access patterns, known failure modes — and CoCo helps the user evaluate options against their own access and each provider's terms. Keys live only in a git-ignored .env (template included, names only), and comment author handles are pseudonymized on ingest.

Checklist

• New directory under samples/ with a kebab-case name
• README.md with what it does, prerequisites, and how to run it
• All source included (7 docs, 5 source dossiers, 5 CoCo skills, 5 SQL files, runnable model)
• Added to the root README under ML and Data Science
• No credentials, tokens, account identifiers, or internal object names
• Verified --sample runs clean from inside the sample directory