Sepp and Lucic give you a formula that reads a trend system's Sharpe ratio off the autocorrelation function and drift of volatility-normalized returns, and on 84 liquid futures it reproduces the realized number with a pooled correlation of 0.99 and a regression slope of 0.96. Worth reproducing, and easy to misread. The 0.99 is an in-sample reconstruction the authors flag as such, not an out-of-sample forecast.
Three systems, one exposure
The paper's first practical claim is that the operational differences between trend systems mostly wash out. A continuous European EWMA filter, an American breakout system with trailing stops, and a sign-based TSMOM rule are constructed from different primitives, yet at matched lookbacks the three deliver annualized Sharpe ratios of 0.47, 0.50, and 0.55 (net of costs and 2/20 fees, 31 Dec 1999 to 30 Jun 2026) against 0.47 for the SG Trend Index. They correlate with the index at 80% on average, and the European and American versions correlate with each other at 95%. Ledoit-Wolf tests do not reject equality of the Sharpe ratios versus the index (p-values 0.96, 0.82, 0.62). The reason the paper focuses on the European system is that its continuous signal admits closed forms; the other two are shown empirically to be slow-span implementations of the same signal, tracked by the European formula at correlations of 0.92 (American, spans above one month) and 0.89 (TSMOM).
Is the alpha just low-frequency autocorrelation?
The attribution runs through a Poisson-kernel reading of the return spectrum. The filter span sets the kernel bandwidth, so short spans read the whole spectrum and long spans read only the mass near zero frequency. The system profits at zero drift if and only if the kernel-weighted spectral mass exceeds one, which the authors phrase as trend-following alpha being excess spectral mass at low frequencies. Longer lookbacks pick up a second term in the squared drift. At a one-year span the trend Sharpe is approximately the squared Sharpe of the underlying instrument. This is a clean specification: the expected-return formula needs only a stationary mean, variance, and autocorrelation function, while the Sharpe ratio adds a linear-process assumption with excess kurtosis entering through a single loading (heavy tails, Student-t with 6 degrees of freedom, lower the pooled gross Sharpe by at most 0.009).
Costs decide what survives
The part an implementer should sit with is the break-even cost logic under AR-1. For a first-lag coefficient of 0.05, gross Sharpe and cost drag both decay as one over the square root of the span, so the break-even cost is nearly span-invariant: it moves only between 37bp and 41bp from a one-week to a two-year span. Realistic costs are 40bp to 60bp. So the short-memory predictability that shows up statistically is not exploitable net of frictions, and whatever net performance exists comes from long memory at one-to-three-month spans. This matters because the lag-1 autocorrelation of normalized returns across the universe fell from about 0.04 in the 1990s to about 0.01 after 2010. The fast-signal trade sits right at the break-even line in the recent regime.
Positive skew without alpha
The skewness result is the cleanest thing in the paper and the most useful for anyone selling crisis-alpha narratives. Under white noise, where the expected return of the system is exactly zero, the aggregated return is still positively skewed at every horizon beyond one day, peaking near half the filter span. Empirically, a single-filter European system at a 100-day span shows aggregated skewness of 2.33 at the 55-day horizon against a closed-form 2.35. Right-skew in trend returns is structural, a property of the product form of the daily return, and does not by itself evidence any predictability. Read that before you attribute a defensive tail to skill.
What we built, and why it printed 0.10
We built a long-or-cash ETF adaptation: 32 US ETFs, weekly close rebalance, EWMA span-33 volatility normalization, the paper's candidate spans plus LS(250,20), picking the filter with positive debiased net Sharpe and sizing at a 15% instrument vol target capped at 10% per name. Over 2015-2025 it returned a Sharpe of 0.10. The 0.10 is ours, from a single automated backtest, and it sits well below the paper's ~0.47 for the European system, same sign and worse magnitude.
Our run is one automated pass, and most of the gap is our construction. We disabled shorting, so we hold a long-or-cash sign-discretized book that behaves more like intermittent equity beta than a two-sided trend system, which strips out the divergence leg that carries much of the paper's alpha. Our universe is roughly 46% equity with 11 S&P sector slices collinear with SPY, against the paper's broader 25% equity across 84 global futures with real agriculture, FX, and short-rate breadth, so we get far less of the cross-asset smoothing behind the portfolio Sharpe. Our 2015-2025 window is exactly the decayed regime where lag-1 autocorrelation sits near 0.01, whereas the paper averages over a longer and more favorable history. And low-vol sleeves like BIL and SHV hit the 10% cap so the vol target is never reached, blunting the sizing. We cannot fully close the gap with what we can see, but every visible difference points the same way, and none of it is evidence against the paper. The formula is an attribution engine that fit 84 futures in-sample; our thin ETF result is a statement about our proxy, not about their identity.
What would change my mind about the tradability, as opposed to the accounting, is an out-of-sample test of the estimated autocorrelation and drift inputs, which the authors leave to future work and which the squared-drift in-sample bias (about a/T, roughly 0.17 in squared annualized drift at the six-year minimum) makes non-trivial. The identity is exact. The forecast is unbuilt.