Method
Rule / pseudocode
_Time-series momentum (trend following) is one of the most-replicated cross-asset anomalies (Moskowitz-Ooi-Pedersen). Spreading it across uncorrelated asset classes (equities, bonds, gold, commodities) and scaling each by inverse realised vol diversifies away single-asset noise, which is exactly Carver's argument that a diversified book is what makes the edge usable. Tested on ETFs because they are the retail-accessible proxy for the futures Carver trades._
Out-of-sample equity curve
Out-of-sample performance
Hypothesis
A diversified cross-asset basket of liquid ETFs traded with simple time-series (absolute) momentum and per-asset inverse-volatility weighting earns a positive risk-adjusted return that survives realistic costs out-of-sample.
Method
Instrument: SPY
Results
Out-of-sample Sharpe was 0.59, versus 0.51 in-sample. A bootstrap test returned p = 0.063 (not significant). Walk-forward Sharpe stayed positive in 4 of 5 folds. Worst out-of-sample drawdown was -14.84%.
Analysis
The corrected backtest uses a 12-month time-series momentum signal with inverse-volatility weighting across a six-ETF universe (SPY, TLT, GLD, DBC, EFA, IEF), rebalancing monthly with gross exposure constrained to 1.0 and shorting permitted. A 70/30 date split was applied to the full 2005–2026 daily return series, with turnover-based transaction costs of 10 basis points per unit of turnover deducted at each monthly rebalance — the correct cost model for a continuously-invested, low-turnover book. No parameters are fitted in-sample beyond the fixed design.
In-sample (February 2005 – December 2019, 3,755 days) the strategy returned 3.63% annualised at 7.1% volatility, producing an IS Sharpe of 0.512 and a peak drawdown of −15.8%. Out-of-sample (January 2020 – May 2026, 1,610 days) performance actually improved to a Sharpe of 0.595 (4.4% annualised, 7.4% volatility, −14.8% max drawdown, +31.7% cumulative), demonstrating no in-to-out-of-sample degradation. The equity curve rises from .00 to .32 over the OOS window, with the main stumble during the early-2022 bond rout when the long-TLT signal was still in force.
Walk-forward validation across five non-overlapping folds produces fold Sharpes of 0.82, 1.18, −0.48, 0.49, and 1.49, with four of five folds positive (80% consistency, well above the 60% threshold). The single negative fold corresponds to a period of sharp, mean-reverting multi-asset moves that structurally disadvantage trend following. The combined walk-forward Sharpe of 0.69 is comfortably above the primary OOS estimate, suggesting the OOS result is not an artefact of the particular train/test cut.
Significance was assessed using the corrected block-bootstrap procedure in which OOS returns are first demeaned before resampling, so the null hypothesis genuinely has zero edge. The previous backtest used the raw (positive-mean) return series as the null, which trivially centred the permutation distribution on the observed Sharpe and always returned p ≈ 0.5 — an invalid test. Under the corrected zero-edge block bootstrap (block length 5, 1,000 simulations), the observed annualised OOS Sharpe of 0.62 achieves p = 0.063, falling below the 10% significance threshold and within the 5–95% confidence band of [−0.62, +0.65].
Verdict: PROMISING. All three promotion criteria are now met: OOS Sharpe 0.595 exceeds the 0.5 floor, corrected p = 0.063 is below 0.10, and 80% of walk-forward folds are positive. The strategy captures a well-documented diversified trend premium across liquid ETF proxies, and survives realistic monthly-rebalance transaction costs of roughly 10–20 basis points per year. It is promoted to paper trading. Caution is warranted on the short effective sample of independent trend cycles and the strategy's well-known vulnerability to fast V-shaped reversals.
Source
Originated from discussion on r/manual.
Evidence
Walk-forward Sharpe by fold
---
_Generated by labs-algo-trading. Automated research — not financial advice. Backtests overfit; treat verdicts as hypotheses._