Episodic Pivots: An AI Blind-Read 1,499 SEC Filings — And Was Most Wrong When Most Confident

A price gap after an SEC filing is arguably the clearest signal the market gives: news lands overnight, and the next morning there is a gap in the chart. The obvious assumption is that the CONTENT of the filing decides the outcome — good news, good trade. This study tests that directly, with a purpose-built experiment: an AI read 1,499 SEC filings completely BLIND, with no knowledge of the price or the outcome, and had to judge from the text alone whether the stock should rise afterward. Across 2,843 trades since 1984, the result is unusually clear — and shows that the CONTENT judgment adds almost nothing to the trade.

Setup: 2,843 trades, 1984 through 2026, survivorship-free

The base is 2,843 Episodic Pivot events (large price gaps, mostly following an SEC filing) from 23 April 1984 through 24 July 2026. The analysis is survivorship-free: delisted stocks stay in the sample, so the result is not just showing the survivors. Costs of 0.2% per leg (buy and sell) apply throughout. The benchmark is a random baseline — the same stocks, but arbitrary rather than actual event days — which averages around -0.43%, the zero line every figure in this study should be read against.

The experiment: an AI reads 1,499 filings blind

The core of this study is a simple but strict experiment: 1,499 of the total events recorded carried an attributable SEC filing. An AI read each of these filings — text ONLY, with no price, no outcome, no information at all about what happened afterward — and delivered a judgment: should this filing push the stock up or not, and how confident is that judgment?

The gap between "the AI expected a positive catalyst" and "it did not" is just 0.29 percentage points — practically no difference. The CONTENT of the filing, read the way an attentive human reader would read it, tells you almost nothing about the actual trade.

The more confident the call, the weaker the trade

More striking than the first finding is the second. Within the cases where the AI judged "should rise," confidence was also recorded, in three tiers (high, medium, low). The ordering runs against intuition.

The more confident the AI was that a filing was bullish, the WORSE the trade actually performed — from +6.64% at low confidence down to just +3.44% at high confidence. A plausible interpretation: whatever obviously sounds good to an attentive reader sounds equally obvious to the market on the gap day itself — and is already priced in before the trade is even possible. Surprise value, not the quality of the news, appears to be what drives the trade.

What actually matters: THAT numbers came out, not WHAT they said

If content barely matters, what does? The answer lies in a much simpler distinction: whether an earnings report was the trigger at all — regardless of whether the numbers were good or bad.

Both the hit rate and the average return sit clearly higher for earnings gaps. The bare fact that a company reported numbers that day — with everything that comes attached: analyst reaction, volume, attention — appears to matter more to the trade than any qualitative read of the filing itself.

Holding longer beats the short exit

Beyond event type, holding period was also tested: a fixed 120-trading-day exit against the short Qullamaggie base exit (an MA20 trailing stop with half the position sold after three trading days). The base group covers 2,843 trades, the 120-day run 2,836 — seven events fall out of scope because of the longer observation window.

The long hold nearly doubles the return of the short exit. Anyone who exits immediately after an Episodic Pivot leaves a substantial share of the move on the table, in this analysis.

The sweet spot: gaps between 30% and 50%

Not every price gap is worth the same. The breakdown by gap size shows a clear sweet spot in the middle size bracket — not among the largest, most spectacular jumps.

Giant gaps above 50% are actually the weakest of all gap classes short-term, at just a 35.3% hit rate and the lowest average return (+3.0%) — likely because this bucket carries a heavier mix of takeovers, bankruptcies, and other special situations with limited further upside. The 30-50% range delivers the best balance of frequency and return.

The "king setup": a 30-50% gap WITH earnings

Combining the gap sweet spot with an earnings catalyst produces the strongest single setup in this study:

Honesty about the sample size is warranted: 52 events in 40 years is rare — barely more than one per year on average. The result is striking but statistically thin, and should not be over-interpreted.

Takeover gaps: essentially worthless as a trade

One result that looks surprising at first glance concerns takeover gaps — price jumps triggered by the announcement of an acquisition.

Takeover gaps average just +0.6% — essentially nothing. The reason lies in the nature of the event itself: in a cash takeover, the price jumps immediately to the offer price and sticks there until the deal closes. There is simply no further move left for a trader to capture. Gaps with no identifiable catalyst — jumps whose trigger could not be clearly attributed — actually performed notably better, at +3.78%.

The framework: portfolio simulation and cost assumptions

A portfolio simulation of the base rule (all Episodic Pivot events) over the full available period since 1993 returns 15.91% per year — versus 10.76% for the S&P 500 total return index (dividends included) over the same period. As a control, a random baseline (the same stocks, but arbitrary rather than actual event days) averages around -0.43% — the zero line every figure in this study should be read against. A cost of 0.2% per leg applies throughout.

Limitations of this study

• Fine-grained subgroups thin out fast and get noisy. The finer the breakdown — gap size × earnings status, for instance — the smaller the sample sizes get. One example: small gaps (10-15%) WITHOUT earnings held up surprisingly well over a long hold (notably better than their short-term result would suggest) — a single finding that looks more like noise than a robust pattern.
• Filing attribution is only reliable from the early 2000s onward. Before roughly 2001-2004, an SEC filing cannot always be cleanly matched to an event and an exact timestamp — the data foundation for the earlier decades is correspondingly patchier than for the more recent past.
• The blind judgments come from a single AI model snapshot at one point in time. A different model, or a later version of the same model, might judge SEC filings differently; the confidence-versus-outcome relationship measured here is a finding of this specific experiment, not a general property of AI judgment.
• Trading costs are a flat assumption. 0.2% per leg approximates bid-ask spread and fees only roughly; for less liquid names the real drag could run higher.

Readers who want to see the same survivorship-free rigor applied to a different pattern can find it in our Qullamaggie Backtest — another study in this series that, among other things, systematically tests short-term gap trading. The Episodic Pivot events analyzed here run on continuously in our Episodic Pivots scanner. More backtest studies live together in Studies.

Figures as of August 7, 2026.

This article is a historical analysis of publicly available price and filing data and not investment advice. It contains no buy or sell recommendation, no forecast, and no statement about any individual company listed today. Past results — whether simulated or real — are not a reliable indicator of future returns. Anyone making investment decisions should assess their own situation and risks, if in doubt with professional advice.