Calibrating Oscillator Signals with Isotonic Regression and a PNN
Summary
The article presents an Expert Advisor signal that turns RSI, Stochastic, and price context into seven directional scores. Each score is bounded around a neutral midpoint, then isotonic regression maps its ranking to an empirical probability. An optional probability neural network (PNN) uses a multidimensional feature state to estimate whether similar historical states were bullish or bearish. The design separates calibration of signal strength from evaluation of the signal’s component pattern, and includes controls for entry probability, a raw-signal gate, and the PNN blend weight.
The article describes a forward comparison in which the PNN-enabled setup reportedly had stronger net profit, profit factor, and Sharpe ratio, alongside more trades and lower drawdown than isotonic calibration alone. However, execution thresholds and stop settings also changed or mattered, so the improvement cannot be attributed to the PNN from this comparison. The author recommends a controlled ablation. The evidence is therefore preliminary, and the supplied excerpt does not establish that the approach generalizes across markets or regimes.
Key ideas
- Seven RSI, Stochastic, and price-context modes generate directional scores on a common scale.
- Isotonic regression calibrates the ordering of raw scores into empirical probabilities.
- A PNN compares a feature state with historical examples and can be blended with the calibrated score.
- Forward results favored the PNN-enabled setup, but execution settings confound attribution.
- A controlled ablation is needed to test whether the PNN itself adds predictive value.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.