Wild day, or quiet?
Nobody can predict which way the market moves — we tested that, and the predictions were noise. But how much it moves follows a real pattern: volatility clusters. Calm follows calm, storms follow storms. So every evening this model publishes one number — the probability that tomorrow the S&P 500 (via SPY) moves 1% or more in either direction — and every morning it grades itself in public.
Read this first: this is a probability about the size of tomorrow's move — never its direction, and never trading advice. It is an experiment with a public record, wins and losses included. What it is useful for is expectations: is tomorrow likely to be a day when headlines move markets, or a day when nothing much happens?
Published after the Tue, Nov 10 close, from data available at that close. Model last refit 2026-11 on 4,234 training days (expanding window from 2010, refit monthly).
The live record
Probationary period: fewer than 100 graded days — treat live scores as preliminary.
| Day | Model | Persistence | VIX | Actual move | Result |
|---|---|---|---|---|---|
| Mon, Nov 9 | 75.2% | 42.6% | 57.3% | ±2.37% | Wild |
| Thu, Nov 5 | 16.5% | 21.7% | 33.9% | ±4.41% | Wild |
| Wed, Nov 4 | 39.2% | 15.4% | 44.6% | ±1.65% | Wild |
| Tue, Nov 3 | 41.9% | 16.0% | 45.3% | ±0.30% | Quiet |
| Mon, Nov 2 | 23.5% | 17.1% | 41.4% | ±0.48% | Quiet |
| Fri, Oct 30 | 25.1% | 11.6% | 35.6% | ±1.77% | Wild |
| Thu, Oct 29 | 22.1% | 11.1% | 34.8% | ±0.38% | Quiet |
| Wed, Oct 28 | 42.3% | 11.1% | 36.5% | ±0.34% | Quiet |
| Tue, Oct 27 | 27.9% | 10.5% | 38.8% | ±0.56% | Quiet |
| Mon, Oct 26 | 33.0% | 10.3% | 41.4% | ±0.71% | Quiet |
| Fri, Oct 23 | 33.4% | 10.6% | 40.4% | ±0.52% | Quiet |
| Thu, Oct 22 | 19.2% | 7.2% | 34.1% | ±1.18% | Wild |
| Wed, Oct 21 | 25.0% | 7.8% | 35.9% | ±0.53% | Quiet |
| Tue, Oct 20 | 30.9% | 7.5% | 38.9% | ±0.34% | Quiet |
| Mon, Oct 19 | 32.9% | 12.9% | 42.7% | ±0.42% | Quiet |
| Fri, Oct 16 | 26.9% | 9.8% | 34.7% | ±1.11% | Wild |
| Thu, Oct 15 | 11.9% | 12.3% | 29.8% | ±0.28% | Quiet |
| Wed, Oct 14 | 13.1% | 9.6% | 29.7% | ±1.08% | Wild |
The backtest (2016–2026)
Before this model was allowed to publish anything, it was tested the way every model here is tested: pre-registered rules, walk-forward — every prediction made only from data available at the time, refit monthly, across 2,708 trading days including the 2020 crash and the 2022 bear market. Wild days were 23.8% of all days. Scores are Brier scores (lower is better); skill is the improvement over each baseline.
| Predictor | Brier score | Model's skill vs it |
|---|---|---|
| This model | 0.146 | — |
| Naive persistence (last 20 days' volatility) | 0.158 | +7.7% |
| VIX-implied (the options market's forecast) | 0.165 | +11.3% |
| Climatology (the long-run average rate) | 0.182 | +19.5% |
The model beat the persistence rule — the hard baseline, since it already captures most of the clustering pattern — in 11 of 11 test years. Its calibration slope was 1.02 (1.00 is perfect): when it said 30%, wild days happened about 30% of the time. That is the whole claim — a modest, honest edge in calibration and accuracy, not a crystal ball.
How it works
A logistic regression — deliberately simple, no black box — reads tonight's market and estimates tomorrow's wild-day probability. Its inputs: realized volatility over the last 5, 20, and 60 days; an exponentially-weighted volatility estimate; yesterday's move and the biggest move of the last week; the VIX level and the gap between the VIX and the 3-month implied volatility; the day of the week; and whether tomorrow is a scheduled Federal Reserve announcement day. Nothing else. No headlines, no sentiment, no price direction.
Data is free and public: SPY prices from Yahoo Finance, the VIX and 3-month implied volatility from the CBOE (the exchange that publishes them; cross-checked against the Fed's FRED database), and the FOMC calendar from the Federal Reserve. The model refits on the first run of each month using every day back to 2010.
Known limitations, stated plainly: only scheduled Fed days are flagged — an unscheduled emergency announcement can't be known in advance, and CPI/jobs-report days are not yet flagged (their historical calendars couldn't be sourced reliably). The model is also, by design, reactive: it can only say "conditions like these have usually meant a wild day," never why tomorrow specifically would be one.
The study behind this page — the pre-registered test design and the full results, including every place the model is weak — is kept with the project and summarized in the backtest section above. If the live record degrades, this page will say so; that is the deal with everything on this site.