Using AI Stock Analysis During Earnings Season Without Fooling Yourself
2026-11-12 ยท hedgewing.ai Research
Earnings season is when AI stock tools are used most and trusted least appropriately. The temptation is to ask the model whether a company will beat, and to read a positive forecast across the report date as an answer. It is not. A model trained on price, volume, and volatility features has no information about the quarter's numbers, and a tool that implies otherwise is overselling. What the model can do is useful, but it is different.
What a price model can and cannot see
Going into a report, the model can see the run-up or sell-off in the weeks before, the compression or expansion of volatility, unusual volume, and how the stock has behaved around past reports. These are real signals about positioning and expectations. What it cannot see is the report itself, which is a discrete jump the price history does not contain. A well-built system responds to a known event inside the horizon by widening its range and pulling confidence toward neutral; the honest reading of a forecast that spans a report date is 'the models do not know, and here is how uncertain they are'.
Before the report
The two weeks before a report are where the short-horizon forecasts are most informative, because they describe the drift into the event rather than the event itself. Check the five-day forecast and the per-model split: unanimous agreement on a pre-earnings drift is a cleaner signal than a split vote. Check the volatility z-score; extreme compression often precedes a large move in either direction, which is a reason to reduce size rather than to guess the direction. And check the implied move from the options market if you have access to it, because that is the market's own estimate of the gap, which the price model is not designed to replace.
On the day
The report day is not a forecasting problem. Any position held through it is a bet on the numbers, and the model's calibrated confidence is explicitly not about the numbers. If you hold through, size for the gap: a position whose worst plausible earnings move would breach your per-trade risk cap is too large, and the 99% VaR figure is a reasonable proxy for 'worst plausible' on an ordinary name, with the caveat that earnings tails are fatter than ordinary tails.
After the report
The weeks after a report are, counter-intuitively, where a price model earns its keep. Post-earnings announcement drift, described at /glossary/post-earnings-announcement-drift, is the well-documented tendency of prices to keep moving in the direction of the surprise for several weeks. The day-one gap is in the price history now, the volatility regime has reset, and the longer-horizon forecasts are once again describing a process the models have seen many times. Compare the post-report forecast with the pre-report one: a model that was neutral going in and leans with the gap coming out is reading the drift.
A routine
Two weeks out: read the five-day forecast and vote split, note the volatility regime, decide on size with the gap in mind. The day before: reduce or hedge anything whose gap risk exceeds your cap. The day after: wait for the nightly rescore, then read the longer-horizon forecasts fresh. hedgewing.ai's analysis pages, for example /analysis/MSFT, used here only to illustrate where the panels live and not as a view on the stock, show the horizons, confidence, per-model breakdown, and VaR together. None of this is a recommendation to trade around any report; it is a framework for not asking a price model a question it cannot answer.