A Master-Model Framework for Regime-Conditioned Price Forecasting: Real Statistical Skill, and Why It Mostly Isn't Alpha

We built a machine-learning system that predicts, for 22 different stocks, funds, and currencies,

roughly what price they will reach weeks to a year from now, using two pieces of publicly available

market-stress information (how risky corporate bonds look relative to safe ones, and how nervous

options markets are about the future versus right now) combined with a purely calendar-based

“same time last year” baseline. For half the instruments tested, that calendar baseline alone is the

single best forecast available — a genuine finding, not a failure, and we treat it as such throughout.

For the other half, the market-stress information adds real, measurable forecasting accuracy. The

harder question is whether “more accurate forecast” means “you can make money trading on it,”

and the honest answer, after testing five different ways of turning the forecasts into trades, is no —

not for any of the 22 instruments, once tested properly. We also asked whether the forecasts’

remaining errors follow any fixable pattern, testing five different correction techniques; for 20 of

the 22 instruments the errors are just noise and every correction attempt made things worse, but

for two — gold and JPMorgan — a genuine, trackable pattern exists and correcting for it makes

the forecasts meaningfully more accurate. That corrected system is now running live, publicly, as

an honest forecast-accuracy dashboard — not a buy/sell signal generator, because we could not

demonstrate that it should be one.