Methodology

For each name we fix a catalyst (an earnings date, a trading update) and an entry price. Several models each return a direction (up / down / flat), a confidence, a rationale, and the single key risk to their view. The ensemble is their aggregate: a direction, a confidence, and a spread that measures how much the models disagree — high spread means low conviction.

Every call ships with a written falsifier: the specific, checkable condition that would prove it wrong. A prediction you cannot lose is not a prediction. This is what makes the log honest — there is no room to reinterpret a miss after the fact.

When a horizon passes we record the realised price and mark each model and the ensemble correct or wrong. The hit rate is measured against a naive baseline (“always up”) — beating coin-flips is easy; beating a baseline that exploits the market’s upward drift is the real bar. Models also carry a calibration score: do their stated confidences match their actual hit rate?

Models do not get an equal say forever. As calls resolve, each model builds a track record, and the ensemble leans on the ones that have earned it — rewarding models whose confidence is calibrated (sure when they turn out right, cautious when they turn out wrong), not just occasionally lucky. Recent form counts for more than a good run long ago, and a model only moves off an equal vote once it has a genuine, sustained record. An unproven model is simply one of the crowd until the log says otherwise.

The panel does not retrain — the underlying models are fixed. What changes is their memory: each model is reminded of how its own recent calls actually played out before it makes the next one, so it can temper a view it has lately been overconfident on. The intelligence is in the feedback loop around the models, not in the models themselves.

A frictionless simulation, purely illustrative: equal stake per call, long on “up”, short on “down”, flat calls sit out. We compare the ensemble’s paper P&L to the baseline’s, plus a confidence-weighted variant. No costs, no slippage, no execution — it is a scorecard, not a strategy.