Most market forecasting asks where price will go. It is the hardest question in finance and the one with the least to show for it. NOVAQ's research suggests a more answerable question with direct practical value: how far?
The difference matters because volatility behaves differently from direction. Quiet periods tend to follow quiet periods and busy ones follow busy ones, a property known as volatility clustering. Direction shows nothing comparable. A system that estimates the size of a session can act on that knowledge even with no view on where the session will end.
The forecast
A pre-open model estimates whether the regular session will be larger than its recent norm. It was tested walk-forward from 2020 to October 2026, on about 1,690 days each for NQ and ES, so each forecast used only information available before that day.
On session size it beat a rolling 250-day base rate in all 7 years (2026 partial), with Brier skill of +0.14 on NQ and +0.17 on ES and an AUC near 0.7. Brier skill shows how much better the probabilities are than the base rate. AUC measures how well days are ranked, with 0.5 as chance, so it does not mean that 70% of forecasts were correct. The model also beat a plain volatility baseline, which matters because a simple average of recent ranges already explains a good part of tomorrow's range.
Calibration
A forecast is useful when its probabilities mean what they say. In the highest forecast band, the session finished above its norm on 93% of days on NQ and 86% on ES. In the lowest band, not one of 298 days reached the largest size category. The 80% ranges covered 81.4% of outcomes on NQ and 77.8% on ES.
Forecast error in points fell by 12.4% on NQ and 16.5% on ES against a simple baseline. The inputs are ordinary market facts known before the open, such as recent volatility, overnight activity and scheduled events. The method used to combine them stays private.
Calibration of the session-size forecast
The same model cannot call direction
Asked to forecast direction, it stayed close to a coin flip, at 0.52 to 0.54 with an interval that includes no skill. It could not separate trend days from rotational days either. This is a finding in its own right: the model has no difficulty with size, and the same information gives it nothing on direction.
What can be forecast before the open: AUC by target
Early volume sharpens the picture
Once the session starts, volume tells more. Projected daily volume can be estimated within about 12% half an hour into the session and about 5% by early afternoon, measured as median error on NQ. Range is harder, moving from about 26% to about 13%.
The slowest fifth of days by early volume pace had about half the remaining range of the fastest fifth, with the correct sign in 28 of 28 comparisons across instruments, checkpoints and years. Pace also separated outcomes inside days already labelled neutral or rotational, so it adds information beyond day type. A simple range model cut error by about a third against a naive baseline.
Remaining range of the day by early volume pace
Volatility carries across sessions
A wide earlier session was followed by a wide opening hour about half the time, against one third by chance. Direction and the order of highs and lows did not carry. Range correlations between sessions also appeared in randomised price paths, which shows that this is clustering of move size and not memory of sequence.
Why it matters
Size is a planning input for stop distance, position sizing, the expected reach of targets and the choice between strategy families. In the recent period, a stop linked to forecast range would have varied risk per trade by about 40% between the quietest and busiest forecast days. Knowing how much the market is likely to move has value even with no opinion on direction.
The finding repeats across the research. Options positioning data helped with size and not direction, and the session research reached a similar conclusion from a different angle.
Limits
The size model is experimental and was not pre-registered. The volume study predates the clean data corpus. As a mechanical filter on strategies, volume pace hurt results, so it serves as context and not as a gate.