Neural Network Price Forecasting with Returns and Wavelet Inputs

Summary

This article outlines a neural-network approach to forecasting financial prices, emphasizing data preparation and input representation. It recommends training on quote changes or log returns rather than raw price levels to reduce dependence on price scale and improve stationarity. It also discusses the tradeoff between a short lag window, which misses older context, and a wide window, which raises input dimensionality; wavelet decomposition is proposed as a way to represent distant history with progressively less detail.

The article describes a Forex forecasting experiment using a hybrid multilayer network and historical quote data, with a later nine-day period reserved for out-of-sample evaluation. It reports a mean squared error of 32 points on that holdout and acknowledges that this is unsuitable for real trading, presenting the result only as an exploratory demonstration. The account gives limited detail about validation, trading costs, or a directional benchmark, so it does not establish predictive or economic value. It also notes that preprocessing choices are central and that network decisions can be difficult to interpret.

Key ideas

• Changes or log returns can be more suitable neural-network inputs than raw quote levels.
• Longer lag windows capture distant history but increase model dimensionality.
• Wavelet decomposition is proposed to compress past price behavior while retaining broad shape.
• The Forex experiment uses a later period for out-of-sample assessment and reports substantial forecast error.
• The example is exploratory and does not demonstrate usable trading performance.

Tags

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