When Two Models Disagree

Two models covering the same stock, one says "buy," the other says "sell." A common instinct is to trust whichever model has been right more often lately, but recent accuracy is a noisy guide over any short window, and switching allegiance every week just chases noise. A more disciplined approach treats disagreement as information in itself, not a nuisance to resolve by picking a favorite.

The standard fix is to combine rather than choose. If both models have a genuine, roughly independent edge, a weighted average of their forecasts — weights based on each model's longer-run track record and how correlated their errors are — usually beats either model alone, because uncorrelated errors partly cancel out. When the two disagree sharply, that disagreement itself is a signal: it often means the position should be sized smaller than either model would suggest on its own, since high disagreement historically correlates with higher realized forecast error.

A second useful check is to ask why they disagree. If one model relies on price momentum and the other on a fundamentals ratio, a sharp split might reflect a genuine regime question (is this a re-rating or a correction?) rather than one model simply being "wrong." Logging the reason for each disagreement, not just the outcome, builds a dataset that eventually tells you which model to trust in which conditions — cheaper information than re-running a full backtest every time the two disagree.