Building a Weighted Oscillator with Best-Worst Preferences and Online Learning

Summary

The Neural Weight Oscillator combines normalized trend, mean-reversion, and momentum readings into a bounded 0–100 indicator. It uses the Best-Worst Method to turn pairwise importance judgments into component weights, then optionally adjusts those weights bar by bar with a small Adam-trained linear model. The indicator also includes an EMA signal line, a histogram around the neutral midpoint, and crossover alerts that require a recent price sweep of a prior high or low. The document explains the component formulas, configurable preferences, and how to interpret the midpoint and overbought or oversold bands.

The article characterizes the learning component as a modest adaptive overlay, not a conventional neural network: its direction target is noisy and it lacks out-of-sample validation. It reports that the fixed weighting core accounts for most of the oscillator’s value, but supplies no independent performance study. The signals and suggested uses are therefore educational or exploratory; users would need to test them on their instruments and timeframes.

Key ideas

• The oscillator blends trend, mean-reversion, and momentum features on a shared bounded scale.
• The Best-Worst Method converts relative importance judgments into normalized component weights.
• An optional online linear model adjusts component influence and slightly shifts the final reading.
• Crossover signals are filtered by extreme-zone conditions and recent liquidity sweeps.
• The adaptive layer has noisy targets and no out-of-sample validation, limiting claims about its predictive value.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.