LSTM Time-Series Forecasting, and an Honest Account of What It Can Do
Sequence modelling on market data, evaluated against a naive baseline rather than against a nice-looking chart. The caveats are the most useful part of this one.
- Client
- Quantitative research project
- Industry
- Financial research
10–25%
Error reduction vs naive
Windowed
Sequence construction
Retrainable
Packaged workflow
Extensible
Additional signals
The problem
The brief was to build a deep learning time-series model for stock price forecasting and, more importantly, to establish whether sequence modelling added anything over the naive baseline on this data.
That framing matters. Financial forecasting demos are notorious for charts where the prediction tracks the price beautifully because the model has effectively learned to output yesterday's value. Any evaluation that does not compare against that baseline is measuring nothing.
What we built
A Python pipeline constructing time-window sequences with strict temporal separation between training and evaluation — no shuffling, no lookahead, no scaler fitted across the boundary, which is the leak that produces most implausible results in this domain.
An LSTM sequence model over those windows, with evaluation checks comparing against the naive persistence baseline at every step rather than at the end.
The workflow is packaged for retraining and extension, so additional signals can be added and re-evaluated without rebuilding the harness.
Results
The model achieved measurable error reduction against the naive baseline — often in the 10–25% range depending on the period evaluated — with improved prediction stability from the windowed sequence construction.
The period-dependence is stated deliberately. Performance varied across market regimes, and reporting a single headline figure would misrepresent what the evaluation showed.
What we'd flag
This is a research artefact, not a trading system. Error reduction against a naive baseline on historical data does not imply profitability, which depends on transaction costs, slippage, position sizing and execution — none of which this addresses.
The most valuable part of the work was the evaluation harness. Temporal separation and a naive baseline are the two things that separate a real result from a chart that looks right, and they are the two things most often skipped.
