Retiring the KNN Forecaster

2026-01-10

Our original trading bot was built on a KNN-based forecaster that learned from a rich set of live market features. We called these snapshots "featuresets" and they included orderbook imbalance, trade slopes, spread, and short-horizon price projections.

That pipeline powered early experiments and backtests, but it was tightly coupled to exchange-specific microstructure. As markets evolved, we wanted a model that generalized better across venues and time regimes with less fragile feature engineering.

The new system uses price-only sequences (OHLC) as input to a tuned Chronos2 transformer. It is simpler, more robust, and easier to operate at scale without a heavy real-time feature stack.

The legacy featureset stream and Firebase/Firestore feeds have been retired. If you need historical exports for research, reach out and we can discuss available archives.