Asset Management / Europe / 2020
Predictive signal models for volatile asset trading
A European fund needed signal models that held up in the conditions where its returns were actually determined. We built an ensemble across market sentiment, technical and alternative data, evaluated specifically under high volatility.
Results
- 6pp
- Improvement in risk-adjusted returns in high-volatility periods
- 19%
- Improvement in position-sizing accuracy
Disciplines
- Quantitative research
- Machine learning
- Trading systems
Context
The fund’s existing signals performed acceptably in aggregate and degraded precisely when they mattered most. Average-case evaluation had concealed this: the periods that determined annual performance were a small fraction of the sample and were being averaged away.
Approach
We changed the evaluation basis before changing the models, so that performance was measured under the volatility regimes the fund actually cared about. That reframing did more for the result than any single modelling decision.
The models themselves combine LSTM and Kolmogorov-Arnold network architectures over market sentiment, technical indicators and alternative data. Ensembling across architectures was a deliberate hedge: the failure modes of the two differ, and in a regime shift it is the correlation of failures that hurts. Position sizing was treated as part of the model rather than a downstream step.
Outcome
Risk-adjusted returns improved by 6 percentage points during high-volatility periods, with position-sizing accuracy up 19%. The evaluation harness was handed over with the models, on the view that the fund’s ability to detect its own signal decay matters more over time than the current generation of models.