Most trading sites publish a hit rate and stop there. This post does the opposite. Below is a full audit of SPY prediction accuracy across the 24 sessions in our rolling public archive — June 30 through August 3, 2026 — including the sessions the model got wrong, the ones it only half got right, and the specific conditions where it stopped working. The headline number is 14 correct out of 24, or 58.3%. The three findings underneath that number are more useful than the number itself.
The Scorecard: 24 Sessions, June 30 to August 3
Each night after the close, the pressure model produces a score. Positive means up pressure into the next open, negative means down. Outcomes are graded against Polygon SPY minute bars using the 09:30–10:05 ET window, and the grading rule is published on our How We Score page: the primary measure is the opening print versus the prior close, with a secondary measure that counts a reading correct if price reverses into the predicted direction inside the first half hour.
| Measure | Result |
|---|---|
| Sessions graded | 24 |
| Correct (published rule) | 14 — 58.3% |
| Correct on the opening print alone | 11 — 45.8% |
| Up-pressure calls | 8 of 12 |
| Down-pressure calls | 6 of 12 |
| Longest losing streak | 1 session |
| Longest winning streak | 4 sessions (Jul 28–31) |
| Median absolute gap | 0.30% |
Finding One: Three of the Fourteen Wins Happened After the Bell, Not At It
This is the least comfortable number in the audit. On the strict measure — did SPY open in the predicted direction relative to the prior close — the model went 11 for 24. That is 45.8%, which is worse than a coin flip. The 58.3% figure only appears once you include the secondary reversal rule, which rescued three sessions: July 22, July 28, and July 31. In each of those, SPY opened against the reading and then traded back through the prior close inside the first thirty minutes.
Is the reversal rule legitimate? It is defensible if your exit is at 10:00 or 10:05 rather than at the bell, because in that case the opening print is not where you get filled. It is not defensible if you exit at 9:31. The honest framing is that the model’s accuracy depends heavily on your exit convention, and any trader using it should know which of those two numbers applies to them. If you close at the open, 45.8% is your number over this stretch, not 58.3%.
Finding Two: Low-Conviction Readings Were Noise
Splitting the 24 sessions by the absolute value of the score produces the clearest signal in the data set.
| Score magnitude | Sessions | Correct | Hit rate |
|---|---|---|---|
| Below 1.0 | 9 | 4 | 44.4% |
| 1.0 to 2.49 | 7 | 5 | 71.4% |
| 2.5 and above | 8 | 5 | 62.5% |
| 1.0 and above (combined) | 15 | 10 | 66.7% |
Nine of the 24 sessions carried a score under 1.0, and those nine went 4-for-9. That is a coin flip with commission and spread attached to it. The 15 sessions with a score of 1.0 or higher went 10-for-15. Same model, same period — the difference is entirely in whether the reading had any conviction behind it.
The practical implication is unglamorous: roughly 38% of nights in this window were nights to sit out. A pressure reading is a decision-support input, not an instruction, and the most valuable thing a weak reading can tell you is that tonight is not the night. That is also the cheapest edge available, because passing on a trade costs nothing.
Finding Three: The Model Was Worst on the Nights That Moved Most
Sorting the same 24 sessions by the size of the overnight gap rather than by score magnitude flips the picture entirely.
- Gaps under 0.25%: 8 correct out of 10
- Gaps of 0.50% or larger: 3 correct out of 8
- Largest misses: July 17 (−1.15%) and July 23 (−1.08%), both against an up reading
The model read quiet nights well and large-move nights badly. That is not a flaw so much as a structural limit, and it follows directly from what the reading measures. A close-of-day pressure score is built from information available at 4 PM ET. It cannot see an 8:30 AM economic release, an overnight headline out of Asia or Europe, or a late guidance revision — and those are precisely the things that produce the biggest gaps. If you want the mechanics of that, our walkthrough of what actually sets SPY’s opening price while you sleep covers the sequence in detail.
A model built on the close is strongest on the nights when nothing new arrives, and weakest on the nights when something does. Any honest reading of this archive has to say so.
What 24 Sessions Can and Cannot Tell You About SPY Prediction Accuracy
Twenty-four observations is a small sample, and it is worth being blunt about what that means. At 14-for-24, the confidence interval around 58.3% is wide enough to comfortably include 50%. None of the sub-splits above — nine sessions here, eight sessions there — comes close to statistical significance on its own. Treat them as hypotheses about where the model works, not as established facts.
Regime matters too. The median absolute gap over this window was 0.30% and the largest was 1.15%. That is a compressed, low-volatility summer stretch. A model’s behavior in a quiet tape says little about how it behaves when realized volatility doubles, and a hit rate measured in July should not be extrapolated into October without checking.
Four Things a Trader Could Take From This
- Know which accuracy number applies to your exit. 45.8% at the bell, 58.3% if you hold to roughly 10:05 ET. Those are different businesses.
- Set a conviction floor before the close, not after. Deciding at 3:55 PM that a 0.4 score is “close enough” is how the 4-for-9 bucket gets traded.
- Check the calendar before sizing. The model’s worst nights clustered around large gaps, which cluster around scheduled events. Size for the gap you cannot control, which is the same reason stop-losses fail overnight.
- Track your own results the same way. Grade every trade against a fixed rule, publish the losses to yourself, and split by conviction. Most traders discover their edge is concentrated in a subset of setups they were diluting with everything else.
Why publish the misses at all
Because a hit rate with no drawdown attached to it is not information. Ten losing sessions in 24, a stretch where wins and losses alternated almost every night, and a coin-flip result on the strict opening measure — that is the actual shape of the tool, and it is the shape you need in order to size positions and set expectations sensibly. Readers who want the historical gap-behavior side of this can also review our earlier look at SPY setups and gap-fill odds around FOMC minutes.
We will run this audit again as the rolling archive turns over. The rules will not change between now and then, which is the only way a scorecard means anything.
This article is for educational and informational purposes only. It does not constitute financial advice, investment recommendations, or a solicitation to buy or sell any securities. Past performance and historical data do not guarantee future results. All trading involves risk, including the potential loss of principal. Always do your own research and consult with a qualified financial advisor before making trading decisions.
