Using TabFM to Classify Raw Candles for Hourly Trading Signals
Summary
This educational FMZ strategy expands recent completed OHLCV candles into fixed tabular features and asks TabFM to classify the next candle as up, down, or flat. Price features are expressed relative to the prior close, while volume is scaled against the window average; the stated construction uses only data available before the target candle. Predictions run at startup and then hourly, while open positions are monitored more frequently.
Trades require sufficient prediction confidence. Position size is tied to account equity and stop distance, with a leverage cap, ATR-based hard stop, and trailing protection after a profit threshold. The document describes a real-model path and mock fallbacks when TabFM is unavailable or fails. It explicitly presents the system as a teaching example, defaults live orders off, and supplies no performance evidence. It also flags mismatched configuration parameters and an unused lookahead setting, so the described defaults may not match the exposed controls.
Key ideas
- Recent raw OHLCV bars are flattened into tabular lag features for next-bar direction classification.
- Price inputs are anchored to the prior close and volume is normalized within the observation window.
- Predictions occur at startup and hourly, while stop management runs on a shorter polling cycle.
- Risk controls include equity-based sizing, a leverage ceiling, ATR stops, and trailing exits.
- The example defaults to signal-only mode and may use mock predictions when the real model is unavailable.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.