Methodology
Point-in-time data, leakage-safe evaluation, simulations with costs, and honest baselines.
Point-in-time discipline
A feature for any date may only use information that was actually available on that date. Macro series arrive with next-day availability and are treated that way in history too.
Purge and embargo
Ten years of daily prices are about 2,500 rows, but 180-day outcomes overlap heavily: there are only around 20 truly independent 180-day periods in a decade. Evaluation removes training rows whose outcomes overlap the test window and adds an embargo gap at least as long as the horizon, so results are not inflated.
Simulations include costs
Historical simulations enter at the next executable price and include fees, spread and slippage on both entry and exit. Contributed capital is always separated from market profit.
Baselines that keep us honest
Every strategy is compared against fixed DCA, keeping reserve, and doing nothing. If an idea cannot beat these simple baselines after costs, it does not reach your plan.
Separate horizons
The 7, 30, 90 and 180-day horizons are analyzed by separate models with separate evidence. A model can be trustworthy at one horizon and still be collecting evidence at another.