EXP-000201
COSMOS — 70/30 split stability of the monthly profile
Hypothesis
The monthly return profile of the first 70% of the backtest (positive-month share, mean monthly return) persists in the final 30%.
Controlled change
No change to the system. Deterministic analysis over the extracted monthly series; decision rule fixed before computation.
Dataset & conditions
Extracted monthly returns, chronological 70/30 split at month 69 of 99.
Before / After
Before
- Period
- 2018-01 → 2023-09
- Months
- 69
- Positive months
- 100%
- Mean monthly return
- 0.16%
- Worst month
- 0.02%
After
- Period
- 2023-10 → 2026-03
- Months
- 30
- Positive months
- 100%
- Mean monthly return
- 0.22%
- Worst month
- 0.11%
Difference
Validation
Out-of-sample, walk-forward, Monte Carlo and forward results are shown only where they exist as data. None exist for this entry.
Decision
Late window kept a positive mean month (0.22%) and its positive-month share (100%) stayed within 10pp of the early window (100%).
AI involvement
- Model
- Claude (Anthropic)
- What AI did
- Authored the split-analysis tooling. The 70/30 split and the decision rule were fixed by a human before any run.
- Human review
- ✓ Reviewed and approved by a human
AI-assisted entries are published only after human review; entries generated by AI without that review are withheld by the build pipeline. AI does not predict markets, and no entry claims otherwise.
Notes
Not out-of-sample validation: the optimisation window of this system is undocumented, so the late 30% may have been seen during tuning. This is a stability check on the recorded equity path.
Backtest improvement does not imply future improvement. The more experiments run, the more likely some succeed by chance — this log exists partly so that multiple-testing risk stays visible instead of hidden.