1. The paper, and the transfer
In July 2026 Kurth, Eisler, Rej and Bouchaud of CFM (Capital Fund Management, a Paris quantitative hedge fund) published "Is Trend Still Your Friend? A Microstructural Account of the Demise of Short-Term Trend-Following". Fast trend went flat around 2009. They reject crowding and blame a change of counterparty: the market makers who warehoused inventory gave way to high-frequency trading (HFT) firms mandated to end the day flat, and those firms will not absorb the persistent directional flow of commodity trading advisors (CTAs). Trend survived where the tick, the minimum price increment an exchange allows, is large relative to volatility, and died where it is small.
This desk works on daily bars, so a mechanism living inside the trading day could be killing my results where I cannot see it. I transferred the test to US large caps in a week, with the fatal caveat registered first: this is a daily-bar equity transfer of intraday futures microstructure, so a clean failure was expected.
| Design | Specification |
|---|---|
| Universe | current US large caps, daily adjusted closes; survivorship biased, so every raw figure is an upper bound |
| Tick regime | a penny for essentially the whole 21-year sample; amended Rule 612 gives tick-constrained names half-penny ticks from November 2025, so the constant-tick assumption holds for the sample period and not beyond it |
| Sorting variable | tick over the product of price and return volatility over the trailing 63 days, ranked monthly across the cross section, split at the median |
| Median annualised volatility, high-ratio tier | 29.3% |
| Median annualised volatility, low-ratio tier | 35.5% |
| Portfolios | decile long-short within tier (long the top tenth by signal, short the bottom tenth), equally weighted, lagged one day, rebalanced monthly |
| Statistics | Newey-West, which keeps correlated returns from overstating the evidence; bar raised for four declared tests |
| Stress rerun | gross halved, costs doubled |
2. What happened
Four tests were declared, continuation and reversal in each tier. Several variants of one idea are several chances to get lucky, so the bar rose to 2.96 on the t-statistic, the measure of distinguishability from luck. All four failed; the best reached 0.10.
| Declared look | NW t | Stressed t | Bar |
|---|---|---|---|
| Continuation, high ratio | −3.40 | −4.86 | 2.96 |
| Continuation, low ratio | −5.72 | −5.95 | 2.96 |
| Reversal, high ratio | −1.33 | −6.95 | 2.96 |
| Reversal, low ratio | 0.10 | −6.39 | 2.96 |
The paper predicts not any single test but the contrast between tiers, continuation relatively stronger where the grid is dense; both legs were already declared, so the contrast costs nothing extra. It replicated in the predicted direction, and the rerun for margin of safety then took it apart.
| The contrast | Value |
|---|---|
| Return | +6.14% a year |
| Sharpe ratio (return per unit of risk) | 0.66 |
| Newey-West t | 3.21 |
| Bar, raised for four declared tests | 2.96 |
| Stressed t, gross halved and costs doubled | 1.64 |
| Base bar | 2.50 |
| Sharpe, in-sample to out-of-sample (data the model never saw) | 0.89 to 0.18, a degradation of 80% |
| Turnover | about 100x a year |
A reader should be able to check the arithmetic rather than trust it. At that turnover the cost line is the size of the gross it is charged against. Cost is not a detail here; it is the position.
| Cost arithmetic | Value |
|---|---|
| House assumption per unit of turnover | 5 basis points, the engine backtester's default |
| Borrow on the short leg while open | 25 basis points annualised |
| Cost at roughly 100 turns a year, base case | about 5% a year |
| Cost under the doubled-cost stress | about 10% a year |
| Gross produced by the contrast | 6.14% a year |
3. Rejected
Rejected. A result that exists only at zero cost is not a strategy. The honest reading is not that the paper is wrong; its mechanism lives at a horizon where my costs are fatal and daily bars record nothing.
The direction replicated days after publication, on a variable never before in my pipeline, then died on turnover, the term a daily-bar cost assumption estimates worst. Signals whose edge sits closest to the trading day are the ones I should trust least.
4. What this does not claim
| What daily bars can test | What they cannot, and I do not pretend otherwise |
|---|---|
| Cross-sectional signals rebalanced daily or slower | Order book dynamics and queue position |
| Conditioning variables built from price, volume, filings and options open interest | Adverse selection at the tick, and the shape of impact within a day |
| Realistic cost sensitivity through a stressed rerun | Anything that requires knowing what happened between the close and the open |
| Whether an effect concentrates in a handful of names or a single regime | Execution: my cost model is basis points on turnover with a doubling stress, a discipline rather than an execution model, and a real slippage model needs tick data and a venue map, which I do not have |
Two further tests are named rather than run, and this note is worth less than it would be if I had run them.
| Not run | What it would settle |
|---|---|
| Slow the tier rebalance from monthly to quarterly | The cheapest falsification: if turnover alone kills the contrast, the slower rebalance should preserve most of the signal and cut most of the cost |
| Regress the contrast on standard factors | Whether this is a new exposure or an old one repackaged. The tier split sorts partly on price and partly on volatility, and short-horizon reversal plus a volatility risk premium would reproduce a good deal of it. Academic for this book, since the result died on cost; load-bearing for anyone who reads the direction as encouraging |
Educational research only. Not investment advice.
Educational research only. Not investment advice.