A rig that makes it-comes-down-to-discretion measurable

The common claim that “discretionary judgment turns a losing strategy into a winner” is often difficult to verify, but it is not impossible. I have implemented a three-part testing framework to quantify exactly how much skill a human trader needs to bridge the gap between a mechanical base and a profitable outcome. My framework uses three core components:

The Bracket Method: This benchmarks a strategy against three tiers: the “Floor” (taking all trades), the “Oracle” (a perfect filter that knows the future), and a “Random Control” (filtering at the same rate as the oracle). If the Oracle isn’t profitable, no amount of human discretion can save the strategy.
The Skill Curve: This models discretion as an “Oracle with noise,” where κ=0 is random guessing and κ=1 is perfect foresight. By calculating the minimum κ* required to reach a target Profit Factor (PF), I can determine if a claimed win rate is statistically realistic.
Blind Measurement: A tool that presents historical data as anonymous charts, stripping away future knowledge. It tracks your decisions to estimate your actual κ and determines if your performance is statistically significant or just luck. I tested this framework on two distinct cases using data from 2015 to 2026 (q=1/3, MC400).

Case A: EmaCrossPullback (USDJPY H4)

This strategy relies on a “coin-flip” mechanical core.

The Numbers: The “Floor” resulted in a PF of 0.78 with a monthly return of -0.03%. The random control returned a PF between 0.40 and 1.34.
The Verdict: To reach a PF of 1.3, a trader would need a κ* of 0.25, which is equivalent to correctly guessing the future 66.7% of the time. Claims of an 80% win rate require a κ* of 0.60, or knowing the future in two out of every three trades. Even with a perfect Oracle, the monthly return is capped at +0.10%. In other words, claiming an 80% win rate is unrealistic because it requires impossible information density; even if it were true, the return is too low to be meaningful.

Case B: Connors Sleeve Core (11 pairs D1)

This is an established mechanical edge (risk 0.6%).

The Numbers: The “Floor” yields a PF of 1.44 with a monthly return of +0.22%. If a trader uses discretion to “selectively filter” trades (q=1/3), the PF stays at 1.45, but the monthly return drops to +0.07%.
The Verdict: Because the frequency of trades decreases, the net return drops to one-third of the original. To recover the original monthly return, the trader needs a skill level of κ≈0.30. In other words, if your discretionary filtering is less accurate than κ=0.30, you are actively degrading a strategy that was already profitable on its own.

Measurement requirements

To distinguish actual skill from random noise, you need a significant sample size of decisions. My data shows that to prove your skill (κ) is not just luck, you need:

κ=0.1: 766 decisions
κ=0.3: 82 decisions
κ=0.5: 28 decisions My testing confirms that this framework is robust. The estimator correctly identifies simulated traders with κ=0.4 and returns no significance for those with κ=0. This means you can now objectively measure your own “discretionary edge.” Before you risk real capital on a manual strategy, use this tool to see if your κ is statistically significant. If you haven’t reached a positive κ, the smartest move is to keep the system purely mechanical.

How this connects

This verification builds on earlier ones (what failed before and what I tried this time, comparisons between approaches).

Code to reproduce

You can reproduce this with the following scripts (see repo).

btengine/discretion.py
scripts/tools/discretion_quiz.py