About EarningsChaser.

EarningsChaser is a daily research publication that uses LLM analysis of fundamentals, market data, and reaction history to predict how stocks will move on earnings prints — and translates each prediction into a concrete options trade with defined risk.

What we look at

For every liquid US-listed name reporting in the next seven days, the system pulls together:

Last 8 quarters of post-earnings 1-day, 5-day, and 30-day reactions (raw + abnormal vs SPY)
Pre-earnings drift and positioning (5d, 14d, 30d) to flag stretched vs washed-out setups
SEC filings — recent 10-K / 10-Q / 8-K and the Form 4 insider transactions in the last 90 days
Analyst conviction trend over 6 months — bullish share rising or falling
News and sentiment from Alpha Vantage and Finnhub, with peer read-across context
Stocktwits + ApeWisdom retail sentiment to flag contrarian setups
Live options chain — implied move, IV skew, ATM IV, put-call activity
Short interest, days to cover, and trend
Market regime — VIX level + 90-day percentile, SPY trend, sector ETF, yield curve

How predictions are made

A 5-step local LLM analyst chain (gpt-oss-20b, run on-prem) processes the dossier — reaction historian, setup reader, sector context, transcript tone analysis, and synthesis — producing a structured prediction with direction, magnitude (always positive), implied move, edge (magnitude minus implied), and calibrated confidence.

Crucially, the model is constrained by override rules: pending binary catalysts force "uncertain" direction; peer read-across dominates "less bad than feared" framing; insider selling at peak prices dampens bullish calls. These guardrails are continuously refined based on misses.

Calibration

Every emailed prediction is scored against the actual reaction. All figures below are measured over every scored event to date, taking the latest belief per event, as of 2026-07-16.

Direction is close to a coin flip: 50.2% at 1 day across n = 2678 scored events. Raising the confidence bar does not help — the 0.60–0.80 band measures 52.6% (n = 447) against 49.7% (n = 2229) below it. We do not treat direction as the signal, and there is no confidence threshold that makes it one.

The tradeable edge is the variance risk premium — whether the realized move is bigger or smaller than the options-implied move. It is computed from features rather than from the language model's opinion, which is why setups are ranked by VRP rather than by directional conviction.

We report it two ways, because they are different claims. On forward calls — predictions made before the print, which is what a subscriber could actually have acted on — it is right 59.6% of the time across n = 47 setups since 2026-05-01, averaging +0.3 vol points. Across backtested events (scored retrospectively at point-in-time cutoffs) it is right 65.1% across n = 447, averaging +1.8 vol points. The forward sample is small and is the weaker of the two; we publish both rather than blend them into one flattering number.

Magnitude validation uses Alpaca historical option chains to compute the implied move at print time, then compares it to the realized move. That comparison is what the VRP figure above scores.

What this is not

Not investment advice. Not personalized recommendations. EarningsChaser is a research publication operating under the SEC's publisher exemption (per SEC v. Lowe, 1985). All subscribers receive the same general-market analysis. We are not a registered investment adviser. Suggested options plays are template strategies — verify strikes against the live chain, sanity-check premiums, never risk more than you can afford to lose to a gap. Past calibration is not a guarantee of future results.

Who's behind this

EarningsChaser is published by Blazing Customs LLC. Questions, feedback, or interested in the paid tier when it launches — admin@blazingcustoms.com.