A Tokenized Transformer Model for 24-Bar Price Direction Forecasts

Summary

The article presents TimeGPT, a compact transformer-style model implemented for MetaTrader 5 to forecast price changes over 24 bars. It describes storing calculations in matrices, normalizing historical price changes, and mapping them into a 256-value token vocabulary. A causal attention mechanism weights each observation against prior observations, with transformer layers used to produce a forecast. The stated design emphasizes running within MT5’s computing and memory constraints, with limited model size and historical inputs.

The text reports 48 trades, an 87% profitable-trade share, a Sharpe ratio of 1.73, a profit factor of 1.49, and 68% directional accuracy. These figures are presented without enough accompanying information here to assess the instrument, period, costs, validation design, or robustness; they should not be treated as evidence of a durable edge. The article also acknowledges that the smaller model may miss long-term behavior and cannot readily incorporate news or sentiment, while larger models require substantially more computing resources.

Key ideas

• The model converts normalized historical price changes into discrete tokens before forecasting.
• Its causal attention layer emphasizes selected past observations when processing a sequence.
• The implementation targets MT5 and uses a relatively compact architecture to limit resource needs.
• The article reports backtest and directional-accuracy figures, but the supplied text omits key validation details.
• The stated trade-offs include limited long-horizon analysis and restricted capacity for adding other data sources.

Tags

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