The story of TyBuff's analytical engines — told honestly, without the recipe.
Most AI trading tools train a model per instrument on that instrument's history. It's the fastest way to look brilliant in a backtest — and the surest way to fail out of it, because the model has memorized one past instead of learning how markets behave. TyBuff's engines were built the opposite way from day one: they never saw a single real stock during their formation. What follows is the shape of that process. The precise techniques, parameters and datasets are trade secrets — but the shape is worth telling, because it explains why the engines behave the way they do.
It started with data science, not price prediction: a sample of equities selected to be statistically and sectorally different from one another — chosen for how differently they behave, not for how well they performed.
From randomized portions of that behavior, we generated hundreds of millions of synthetic price histories — markets that never existed, but statistically could have. This is the step that removes memorization: there is no real ticker to overfit to.
Machine-learning clustering and systematic cross-validation searched that synthetic universe for coherent families of market behavior — patterns that persist because of how markets work, not because of what one stock did in one decade.
The process converged on fifteen distinct, internally coherent analytical logics. Each became an engine. Not one was designed by hand around a market story — each one earned its existence by surviving the clustering and validation process.
Only then did the engines meet real market data — the full history, none of it seen during formation. Each logic had to prove its coherence there before earning a place in the platform. The ones you use today are the ones that did.
We publish the shape of the process — never the recipe.
The specific techniques, parameters, datasets and validation thresholds are proprietary trade secrets and are not disclosed.
Because no real instrument was memorized, the same engines apply one consistent analytical lens across stocks, ETFs and crypto — no per-asset retraining, no per-asset excuses.
Overfitting is usually fought with discipline after the fact. Here it was designed out at the source: you cannot overfit to a specific history you never trained on.
Engine behavior is version-controlled and frozen. Any change to an engine must reproduce its previously verified outputs exactly — to the last decimal — before it ships. Your saved strategies stay reproducible.
The engines evaluate universes built on point-in-time data: at every historical date, only the instruments that actually existed then — including ones that were later delisted. No survivorship bias, no hindsight.