Using TabFM to Classify Next-Bar Direction from OHLCV Windows

Summary

This article outlines a way to use the TabFM tabular foundation model for short-horizon market prediction. It converts completed OHLCV candles into fixed-length lagged rows, scales prices relative to the latest close in each window, and scales volume against the window average. Historical rows are labeled as up, flat, or down according to the next candle’s close-to-close return and a chosen threshold; the current completed window is then classified using those examples as context.

The method aims to let a pretrained tabular model discover relationships in price and volume windows without manually adding indicators such as moving averages or RSI. The article explains sample alignment, exclusion of unfinished candles, and the distinction between this tabular approach and a dedicated time-series model. It presents no sample-out-of-sample performance evidence and explicitly treats the work as a methodological experiment. Thresholds, window length, costs, inference speed, and model licensing all limit practical use and require further evaluation.

Key ideas

• Represent each sample with lagged OHLCV values from completed candles and label it using the following candle’s direction.
• Normalize prices to the most recent close in each window and scale volume relative to that window’s average.
• Choose a return threshold carefully because it affects both label noise and the balance among classes.
• TabFM receives a fixed-column table, so lag positions encode temporal order rather than the model handling a native time series.
• The described experiment does not establish predictive or trading performance and needs out-of-sample and cost-aware validation.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.