RSI, Feynman Principles, and Epistemic Methods
Overview
Exploration of Monte Carlo methods as epistemic instruments, RSI (Relative Strength Index) as an AI training feature, and the Feynman Principle (“The first principle is that you must not fool yourself”) as a guardrail for financial AI.
Monte Carlo as Epistemic Instrument
MC methods transform probabilistic sampling into justified beliefs:
- Quantifying epistemic uncertainty — Modeling “unknown unknowns” in limited data scenarios
- Simulation as inference — Deriving knowledge through logical inference, not just raw data
- Sensitivity analysis — Separating epistemic uncertainty (knowledge gaps) from aleatory uncertainty (randomness)
RSI in AI: Beyond Simple Thresholds
Regime Detection (Context Layer)
RSI oscillating 40-80 indicates Bullish regime; 20-60 indicates Bearish. Train a classifier to adjust risk parameters based on this “background radiation.”
Divergence Detection (Alpha Layer)
Price makes new high but RSI makes lower high — “the price is fooling the market, but momentum tells the truth.” Train CNNs/LSTMs on RSI curve shapes for automatic divergence detection.
Failure Swings
Identify RSI breaks of previous low/high as precursors to trend reversals. Effective for optimization of trade exit timing.
Training Methodology: Adding Noise
Why Add Noise
Forces AI to learn data structure rather than memorize specific points (Denoising Autoencoding / Domain Randomization).
| Noise Type | Method | Purpose |
|---|---|---|
| Gaussian (White) | Inject N(0, 0.01) into RSI inputs | Simulate market microstructure noise |
| Adversarial | Intentionally alter data at worst points | Train model stubbornness against single-point flip-flops |
Detection vs. Optimization Workflow
- Detection — Train with heavy noise; robust detectors find signals through static
- Optimization — Once detected, use cleaner data to refine entry/exit timing
The Feynman Principle: Anti-Fooling Guardrails
The “Luck” Test
Run MC simulation on shuffled price returns (destroying trend, keeping volatility). If the RSI strategy finds profitable trades in nonsense data, you are detecting noise, not signal.
Devil’s Advocate Module
If AI says “Buy,” automatically search for the strongest counter-argument (earnings reports, low liquidity, etc.).
Anti-P-Hacking
If you tweak Gaussian noise parameters until the model “works,” you are cheating. The model must work on distributions it hasn’t seen.
Feynman Stress Test
Take a signal (e.g., RSI < 30), run 1,000 MC simulations of next hour using current-volatility-derived Gaussian noise. Only pass to user if signal survives 95% of noisy futures.
Related
- monte-carlo-simulation-for-financial-forecasting — Financial Monte Carlo simulation research
- synthetic-data-generation — Synthetic data generation patterns