Machine learning on tick data: separating hype from reality

[Shadow Trader]

Machine learning in trading has a glamorous image. Feed in tick data, let a neural network find patterns, collect profits. The reality is harsher.

Financial data has a very low signal-to-noise ratio. Models easily fit noise, producing gorgeous backtests that collapse in live trading. Markets are also non-stationary, meaning patterns shift over time, so a model trained on last year's behaviour may have little relevance next year.

Costs are the other killer. A model predicting 0.3 pips of movement over the next few seconds may be statistically significant and still unprofitable after a half-pip spread.

That doesn't mean machine learning is useless. Simple models used carefully, with proper validation, honest cost modelling and strict out-of-sample testing, can add value, especially for classification tasks like regime detection or spread prediction.

The red flags are familiar: tiny test periods, no costs, tuning on the test set, and claims of astonishing accuracy.

What has been tried with machine learning, and did anything survive realistic testing?