Prometheus Basic Trend: publisher stats match, holdout edge does not
Completed 1319 · 2026-08-27
Can QuantModel reverse-engineer the Prometheus Macro Basic Trend program from public posts so that published performance stats match, and should that replica enter Other / Strategies?
The causal public-data replica clears every publisher fidelity tolerance on the approximate 2001–2026 window (Sharpe 1.445 vs 1.44). It does not clear the frozen untouched 2023–2025 economic floors versus its own static beta (Sharpe improvement −0.64 vs +0.20 required). Disposition is fidelity_only: keep the implementation and evidence, publish the research surfaces, and do not add a strategy-catalog row.
Provenance
finding · artifact · code · payload · notebook analysis
Research narrative
The short version
We rebuilt the disclosed Basic Trend architecture on public data. The long-run published stats line up. The 2023–2025 holdout does not beat the static beta enough to admit a catalog strategy.
What was frozen
21/126 trends, 126-session inverse vol, 21-session covariance, 3x cap, next_close lag, 10 bps costs, SPY/TLT←VUSTX/GLD←GC=F/Coin Metrics BTC, and numeric promotion floors were fixed before confirmatory returns.
What failed
Untouched Sharpe improvement −0.64, MDD improvement only +3.4pp, adjacent parameters 0/5, extra lag and 2x costs stay negative versus beta.
Why no catalog entry
research.strategy_evidence.admission admits only promote. This run is fidelity_only, so the public page is an experiment/result, not Other / Strategies.
Basic Trend replica vs static beta — annual growth of $1
Aligned window, both rebased to 1.0 at start — Both curves use base 10 bps one-way costs and next_close execution on the frozen public panel through 2026-07-24. · S&P dashed, experiment solid. X = Date, Y = Equity (rebased to 1.0).
Signal driver & thresholds (x-aligned)
Documented vs inferred
| piece | documented | hidden |
|---|---|---|
| Universe | Stocks, bonds, gold, bitcoin sleeves | Exact investable tickers |
| Trend | 1m and 6m full/half/flat multipliers | Exact session counts until 21/126 inference |
| Sizing | Inverse-vol base + breadth-scaled 0–15% vol target | 126/21 estimators via allocation MAE |
| Leverage | Long only; residual cash | 3x gross cap and T-bill financing are QuantModel choices |
| Execution | Daily signal histories published | next_close lag and 10 bps costs are repository assumptions |
Inferred values were frozen before return evaluation and were not retuned afterward.
Untouched window stresses (Sharpe vs beta)
| check | sharpe_improvement | mdd_improvement | result |
|---|---|---|---|
| RET-PRIMARY | -0.639 | 0.034 | fails floor |
| RET-GROSS | -0.355 | 0.050 | fails floor |
| RET-COST2X | -0.923 | 0.017 | fails floor |
| RET-LAG2 | -0.541 | 0.055 | fails floor |
| RET-TREND-20-120 | -0.665 | 0.048 | fails floor |
| RET-TREND-22-132 | -0.443 | 0.058 | fails floor |
| RET-VOL-63 | -0.619 | 0.036 | fails floor |
| RET-VOL-252 | -0.647 | 0.028 | fails floor |
| RET-COV-42 | -0.581 | 0.042 | fails floor |
| RET-NATIVE | -0.182 | 0.059 | fails floor |
Primary promotion floors require Sharpe improvement ≥ 0.20 and MDD improvement ≥ 0.08. No failed cell was tuned away.
Stats
| CAGR | 16.35% | Sharpe | 94.02% |
|---|---|---|---|
| Sortino | 135.58% | maxDD | -9.84% |
Disposition
Fidelity only — discoverable experiment and results pages only; no Other / Strategies catalog admission under research.strategy_evidence.admission.
Follow-on work
- Catalog revisit if economic floors later clear
Errors & data gaps
Publisher estimators, instruments, costs, and execution timing remain QuantModel inferences selected by allocation MAE before returns. Untouched 2023–2025 fails frozen Sharpe/drawdown promotion floors despite full-period publisher-stat fidelity. BTC PriceUSD is a non-executable reference; capacity and custody are unmodeled. High turnover and leverage up to 3x make net results cost-sensitive. Issue acceptance asked for Other / Strategies publication; fidelity_only forbids catalog admission.
Follow-up
gh issue create --title "Follow-up Prometheus Basic Trend publisher stats match holdout edge do" --body "Follow-up to 1319: Can QuantModel reverse-engineer the Prometheus Macro Basic Trend program from pu" --label "priority:p2" --label "pipeline:research" --label "agent:ready"