Adaptive Temporal Fusion Transformers for Crypto Price Forecasting

Summary

The paper proposes an adaptive Temporal Fusion Transformer approach for short-term cryptocurrency price forecasting. It addresses non-stationarity and sharp price variation by segmenting the series into variable-length subseries that end at relative maxima: points where gains from a preceding minimum exceed a threshold. This is intended to mark meaningful upward phases while filtering some noise. A fixed-length pattern at the end of one segment determines the category of the next segment, grouping subsequent price paths by their preceding conditions.

A separate TFT is trained for each category to forecast the evolution of the next segment from its initial observations. Experiments use ETH-USDT ten-minute data and a two-month test period. The document reports better prediction accuracy and simulated trading profitability than fixed-length TFT and LSTM baselines. Those results are specific to the stated asset, interval, and test period; the document does not establish whether the method transfers to other cryptocurrencies, market regimes, or live trading after execution costs.

Key ideas

• Variable-length segments end when gains from a preceding minimum exceed a threshold.
• The final fixed-length pattern in one segment assigns a category to the next segment.
• Separate TFT models forecast subsequent segments within each pattern category.
• The reported ETH-USDT test outperforms fixed-length TFT and LSTM baselines in accuracy and simulated profitability.
• The evidence is limited to the stated asset, sampling interval, and test period.

Tags

# Adaptive Temporal Fusion Transformers for Cryptocurrency Price Prediction


# Adaptive Temporal Fusion Transformers for Cryptocurrency Price Prediction









Precise short-term price prediction in the highly volatile cryptocurrency market is critical for informed trading strategies. Although Temporal Fusion Transformers (TFTs) have shown potential, their direct use often struggles in the face of the market's non-stationary nature and extreme volatility. This paper introduces an adaptive TFT modeling approach leveraging dynamic subseries lengths and pattern-based categorization to enhance short-term forecasting. We propose a novel segmentation method where subseries end at relative maxima, identified when the price increase from the preceding minimum surpasses a threshold, thus capturing significant upward movements, which act as key markers for the end of a growth phase, while potentially filtering the noise. Crucially, the fixed-length pattern ending each subseries determines the category assigned to the subsequent variable-length subseries, grouping typical market responses that follow similar preceding conditions. A distinct TFT model trained for each category is specialized in predicting the evolution of these subsequent subseries based on their initial steps after the preceding peak. Experimental results on ETH-USDT 10-minute data over a two-month test period demonstrate that our adaptive approach significantly outperforms baseline fixed-length TFT and LSTM models in prediction accuracy and simulated trading profitability. Our combination of adaptive segmentation and pattern-conditioned forecasting enables more robust and responsive cryptocurrency price prediction.

Shown in full with attribution under the source's licence. Licence: abstract CC0

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