Using Daily Price Normalization and Time Tokens for LSTM Forecasts

Summary

The article presents a preprocessing method for training an LSTM to forecast intraday currency prices. It encodes clock time as a fraction of the day and scales each day’s open, high, low, and close relative to the expanding intraday low and high. These rolling bounds reset at the start of each day, placing prices on a comparable scale while retaining their position within that day’s observed range. The model uses these normalized OHLC features together with the time token, and its output is converted back to a price using a selected daily range.

The author describes a 15-minute EURUSD dataset, model inputs, and visual checks of normalized values, but the supplied text does not establish predictive or trading performance. The author notes that reconstructing earlier predictions with a full day’s high and low is inaccurate because those bounds were not yet known at each point in time; a rolling range would be needed. Data availability, demo-account spread quality, and omitted volume and spread features also limit the setup. The method is an exploratory forecasting workflow, not evidence of a validated trading edge.

Key ideas

• Time of day is represented as elapsed seconds divided by the seconds in a day.
• Intraday OHLC prices are scaled against expanding daily highs and lows that reset each day.
• The LSTM uses time and normalized OHLC values as its input features to predict normalized price.
• Converting predictions back to historical prices requires range information available at the prediction time.
• The article shows preprocessing and visualization but provides no demonstrated forecast accuracy or strategy returns.

Tags

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