How to Read a Per-Model Vote Split (and When to Ignore the Headline Forecast)
2026-11-12 ยท hedgewing.ai Research
When an ensemble forecast says 61% probability of a rise over ten days, it is tempting to stop reading. But that single number compresses four separate opinions, and the way they were distributed is information you should not throw away. Two forecasts at 61% can come from four models that all lean mildly positive, or from two models that are strongly positive and two that are negative. The first is a consensus; the second is a disagreement that happened to net out.
What each model is good at
The split is more readable when you know what the voters are. The LSTM and GRU are recurrent networks that read the price sequence step by step and tend to pick up momentum and mean-reversion over weeks. The temporal convolutional network sees multi-scale patterns through dilated convolutions and is often first to notice a structural change. The Transformer relates any two days in the lookback window directly and is most useful when the relevant context is not recent. Because their inductive biases differ, agreement among them is more meaningful than agreement among four copies of the same architecture.
Reading the common patterns
A unanimous 4-0 vote with moderate confidence is the cleanest signal the system produces: different models reached the same conclusion from different angles. A 3-1 vote is normal and still usable, but it is worth looking at which model dissented; if the dissenter is the one that has been most accurate on this stock recently, weight the split more heavily. A 2-2 vote means the ensemble's headline is being decided by the meta-learner's weights rather than by broad agreement, and the honest reading is 'no edge', regardless of which way the combined number leans. A vote where the recurrent models and the convolutional model point in opposite directions often marks an inflection, where the trend is intact in the sequence view but the pattern view sees it breaking.
Combining the split with confidence
Confidence and agreement are related but not identical. High confidence with a split vote happens when one or two models are extremely sure and the others mildly opposed; treat it as fragile. Moderate confidence with a unanimous vote is often the more actionable combination because it is robust to any single model being wrong. A simple rule is to require both a confidence threshold and at least a 3-1 vote before sizing a position near their risk cap, and to treat 2-2 as an instruction to wait for the next nightly update.
When to ignore the headline entirely
There are situations where the split should override the number. If a known earnings date falls inside the forecast horizon, all four models are forecasting across a gap event none of them can see, and agreement is less reassuring than it looks. If volatility is in an extreme regime, the models are extrapolating from conditions they have seen rarely. And if the split has flipped sign on consecutive days, the signal is unstable and any single reading is noise. In each case the right response is smaller size or no position, not a cleverer interpretation.
Where to find it
On hedgewing.ai, the per-model breakdown is shown alongside the ensemble forecast for Pro users on each stock's analysis page, for example /analysis/AAPL, used here only as an illustration of where the panel appears and not as a comment on the stock. The glossary entry at /glossary/ensemble-disagreement explains the measurement. Nothing in this article is a recommendation to buy or sell any security; model outputs are research inputs, and backtests on the platform are in-sample and illustrative.