Using TabFM to Classify the Next Candlestick from Raw OHLCV Windows

Summary

This article describes a quantitative prediction workflow that turns each recent candlestick window into a tabular sample for TabFM, a foundation model for tabular data. Each row contains OHLCV values from completed bars, arranged as lagged fields, and the label classifies the following bar as up, down, or flat. Prices are normalized relative to the latest close in the input window, while volume is scaled by the window's average, to reduce dependence on changing price levels. The method leaves feature discovery to the model rather than computing indicators such as moving averages or RSI in advance. The researcher still chooses the observation window, timeframe, sample set, and return threshold, and must ensure that inputs contain no future data. The document explicitly frames the example as an experiment: it reports no complete out-of-sample test or systematic comparison with other models, so it does not establish predictive or trading value. Parameter choices and validation remain essential.

Key ideas

• A rolling window of completed OHLCV bars becomes one tabular input row with lag fields.
• The target label assigns the next bar to up, down, or flat based on a chosen return threshold.
• Price features are normalized to the latest close, and volume is scaled by its window average.
• TabFM may learn relationships from raw windows without manually specified technical indicators.
• Window design, labels, data selection, and out-of-sample validation remain researcher responsibilities.

Tags

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