Fair questions, straight answers

A nightly pipeline that scrapes market data, warehouses it, and trains models to answer one auditable question per stock: how likely is a +5% move within the next five sessions? The output is a probability — never a point forecast — and every probability is later graded against what actually happened.

No. Everything on this site is research output from a personal project, published for transparency and education. Probabilities are graded honestly — including the misses — and nothing here is a recommendation to buy or sell anything.

Access is limited while the platform is tested end to end. Public access will open once the nightly pipeline and its grading have been validated over a longer live stretch.

Early users will be able to explore for free. Every additional user adds server, data, and compute cost, so some form of monetization is likely later — the goal is for any price to be small next to the value of the signals themselves.

Three pillars: SEC EDGAR XBRL filings (fundamentals with exact publication timestamps), a market-data API covering prices, corporate actions, share counts, earnings, analyst estimates, news sentiment, insider transactions and 13F institutional holdings, and FRED macro series. Independent sources are reconciled against each other — a standing data-quality alarm compares SEC and vendor fundamentals every night.

Daily prices reach back roughly 26 years and insider transactions to 2004. SEC XBRL fundamentals begin around 2009 (when structured filing became mandatory). News sentiment and analyst-estimate vintages accrue from mid-2026 onward — those histories cannot be backfilled, which is exactly why capturing them daily matters.

Every input joins the training data only after the moment it became public knowledge: filings at their acceptance timestamp, earnings at their publish time (pre- vs post-market), insider trades after the two-day Form-4 deadline, 13F aggregates after the 45-day filing deadline. Warehouse tests enforce this on every build, so backtests can't cheat and live results match tested ones.

Barrier-hit probabilities. A '+5% / 5d' signal is the probability that the stock's high crosses +5% of the next session's open within five trading days — barrier rounded to the cent, entry at the next open. Threshold questions with gradeable answers, not price targets.

Gradient-boosted decision trees (CatBoost classifiers), retrained on a rolling window in a walk-forward sequence — each retrain only ever sees data available at its train date. Simple, fast to retrain nightly, and honest about feature importance.

Three ways, all published: out-of-time AUC for every retrain (scored on dates the model never saw), daily realized hit rates of the model's top picks against everything it scored, and calibration by prediction decile — do stocks given a 20% probability actually hit about 20% of the time? The grades include the misses.