Testing GARCH Volatility Filters with a Volatility-Scaled LSTM
Summary
The article describes an Expert Advisor signal that combines price-action patterns with a GARCH(1,1) estimate of near-term volatility expansion. Seven signal modes interpret breakouts, squeeze releases, band re-entries, impulses, band walks, pullbacks, and range escapes using ATR, Bollinger Bands, and price behavior. The LSTM option processes short sequences of normalized volatility features and adjusts the entry likelihood, while sharing the same directional logic and GARCH filter as the simpler configuration.
The experiment compares the deterministic signal with and without the recurrent network, using backtests and a forward walk. The reported results do not show a dependable improvement from adding the LSTM; in one forward run it filtered out a profitable short trade that helped the baseline. The author suggests that the network's features may overlap with information already present in the indicators and patterns. These findings are specific to the tested configurations and do not establish that recurrent models are generally unsuitable for volatility forecasting. The article recommends keeping model components comparable and removable, and requiring clear out-of-sample gains before accepting added complexity.
Key ideas
- GARCH estimates whether forecast volatility is expanding relative to its recent baseline.
- Seven pattern modes combine volatility measures, Bollinger Bands, and price action to identify candidate entries.
- The LSTM supplements the shared signal logic rather than acting as an independent strategy.
- Backtests and a forward walk did not establish a reliable advantage for the network-enhanced version.
- Overlapping inputs may leave the LSTM with little new information to learn.
- Added model complexity should be retained only when it demonstrates out-of-sample value.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.