⚠️ Personal research and trading journal — not investment advice. The author does not provide licensed advisory services.
In May 2026 I separated my paper trading into three independent systems. They run in parallel, with independent capital allocation, through November 2026. At that point an arbiter evaluates the walk-forward results and selects which approach gets the live capital.
This is unusual. Most traders optimize one system. I'm deliberately running three competing ones. Here's why.
The Problem With Optimizing One System
When you refine a single system over years of testing, you create a survivor that has been tuned to your historical data. The refinements that made it "better" may be responses to noise that won't repeat. This is the standard overfitting problem, and it's well-known.
The less obvious problem: you lose the ability to tell whether the system is generating alpha or whether you've been in a favorable regime that masked its weaknesses. A momentum system that was built and tested in 2012-2021 might look impressive not because the methodology is sound, but because the US market had an unusually long bull run.
The conventional fix is walk-forward testing with out-of-sample periods. I use that. But there's an additional question: does the specific implementation of a methodology outperform alternative implementations of equally valid methodologies?
The Three Branches
Pattern branch (current): Chart-based momentum. Breakout from contracting base, RS≥80, Stage 2, volume confirmation at entry. Exit via partial-TP at 2R + MA21 trail. This is the longest-running branch — the 20-year walk-forward validation is on this system.
O'Neil branch: CANSLIM-adjacent. Same entry criteria as pattern branch, but layered with earnings acceleration (C), annual EPS growth (A), and eight O'Neil sell rules (20-25% profit + EMA50 trail, 7-8% hard stop, climax exit, distribution day sell). Different exit structure, different position-duration behavior.
Simons branch: Pure quantitative. Five-factor equal-weight ensemble — RS rating, price-to-52-week-high, earnings growth slope, volume trend, cross-sectional sector momentum. Rebalances weekly. No chart reading. No human judgment on entries.
Each branch runs with ฿333,000 in paper capital. They trade the same universe (US + Thai SET50/100) but apply independent entry and exit logic.
How the Arbiter Works
The branch arbiter, implemented in November 2026, evaluates each branch on four statistical criteria:
1. Per-trade profit factor: gross wins ÷ gross losses. Needs to exceed 1.0 with margin; values below 1.3 are marginal.
2. Deflated Sharpe Ratio (DSR): Sharpe corrected for the number of trials. The standard Sharpe inflates with more tests — DSR adjusts for how many configurations were tried before arriving at the current one.
3. Probability of Backtest Overfitting (PBO): cross-validation test that checks whether the system that performed best in-sample performed best out-of-sample in at least half of re-sampled CV splits. A high PBO (>50%) means the system probably overfit.
4. Bootstrap confidence intervals: nonparametric CI on the 30-day forward return distribution. Must exclude zero with acceptable width.
A branch needs to pass on at least 3 of 4 criteria to be considered. If multiple pass, the one with the highest DSR × (1 - PBO) composite score wins. If none pass, the live allocation stays in cash until one does.
Why Three Rather Than Two
Two branches would leave open the question of whether the winner won due to methodology or regime. Three branches that are methodologically distinct — discretionary chart, CANSLIM hybrid, pure quant — create a more informative competition. If all three fail, the arbiter is telling me something about market conditions, not just about my implementation choices. If one consistently outperforms across varying conditions, the evidence is stronger.
The branches are designed to have different correlation structures. Pattern and O'Neil share entry logic but differ on exit. Simons is independent on entry. This means: - A period where Pattern and O'Neil both outperform Simons tells me the chart read adds value - A period where Simons outperforms both tells me the quant signal has edge the chart read doesn't - A period where O'Neil outperforms Pattern tells me the exit structure matters more than which signals I already know are correlated
What I Know From Prior Replay Analysis
In a replay study conducted in June 2026, I applied each branch's exit rules to the same set of paper entries (taking the intersection of trades that all three branches would have entered), and computed realized R-multiples.
Result across the test set: - Pattern branch: +1R median - O'Neil branch: -1R median - IBD simulation: -17.4R median - Current paper exits: -9.3R median
The spread of 18.4R from exit choice alone — on the same entries — suggests the exit framework is the primary driver of system performance in this data set. That's the hypothesis the November arbiter will test with live-data evidence.
What Happens After November
One branch takes the live capital. The others don't disappear — they continue running as paper, generating ongoing data. If the live branch's performance diverges from the paper continuation of the other branches in a way that's statistically significant, that becomes evidence for reconsidering the allocation.
The system is designed for ongoing evaluation, not a single winner-take-all decision that's never revisited.
Track. Study. Wait. Strike.
Personal research and trading journal — not investment advice. The author does not provide licensed advisory services. — MOEasymmetry
Draft 2026-06-12. Source: 3-branch architecture implemented May-June 2026. Branches: simons/oneil/current (pattern). Each ฿333K paper. Branch arbiter: DSR + PBO + bootstrap CI, November 2026. Replay study: same entries × 4 exit rules → Pattern +1R / O'Neil -1R / IBD sim -17.4R / Current -9.3R (June 2026). Full arbiter methodology at vault [[Branch-Arbiter-DSR-PBO-2026-06-10]] and project_3_branch_architecture.md.