Field Note 07 / Trend Following

A Trading Edge That Wasn't Wrong, Just Old

Of the four hypotheses this series set out to test, time-series momentum—trend-following—came in with the strongest résumé by far.

Moskowitz, Ooi & Pedersen tested it across 58 liquid futures and forward contracts spanning equity indices, currencies, commodities, and government bonds, over more than two decades of data. The effect showed up in nearly every asset class they checked. That's a rare thing in finance: a simple mechanism—an asset that's been rising keeps rising, on average, for a while longer—confirmed across markets that have almost nothing else in common. Trend-following managed futures funds have existed as a real segment of the asset management industry for decades on the back of exactly this result. Coming into this series, trend was the candidate expected to survive the crypto test most easily.

A strategy with real pedigree, tested on the wrong decade

Liu & Tsyvinski found a strong time-series momentum effect for Bitcoin too, with statistically significant results across one-to-four-week horizons. But their sample runs only through May 2018. That's a market with no deep Binance perpetual futures, minimal institutional participation, and none of the structural shifts crypto has been through since: the 2020–2021 bull cycle, the 2022 collapses, and a derivatives market that matured considerably in between. A strong result on 2018 data is real evidence. It's evidence about a market regime, not a permanent property of crypto.

The decisive test came from Grobys, Kolari, Sandretto, Shahzad & Äijö in 2025, using 30 large-cap coins over 416 weekly observations from January 2016 through December 2023, the longest and most methodologically careful crypto-specific test available. Split by regime, the picture flips. Pre-July 2020: 1.74% per week, significant only at the 10% level. Post-July 2020, the regime closest to today's market: negative, and not statistically significant.

One outlier, most of the story

Part of why the raw full-sample result looks weak traces to a single event. On December 23, 2020, the token Mindol rose from $0.54 to $7.86 in a day, a roughly 1,400% move. Sitting in the short leg of a six-coin portfolio, that one move produced a single week's loss of about −255% for the strategy. Trim that one outlier and the full-sample result recovers to a significant 1.51% per week. A volatility-managed version of the strategy, using an 8-week rolling window, reaches 1.86% per week at 5% significance.

Neither number rescues the strategy for right now. Both are full-sample figures, weighted toward the older, more favorable pre-2020 regime. Neither one independently shows the effect holds specifically in the post-July-2020 data, which is the only regime a trader today can actually trade in.

The universe you can't actually trade anyway

The paper also runs a broader check across 2,500 coins, value-weighted. There the result is worse, not better: −3.40% per week, significant, and it stays negative through every trimming approach tested. Separately from whether that number is real, it was never a reproducible retail portfolio to begin with. No exchange, then or now, has listed anything close to 2,500 coins with tradable perpetual futures depth on both the long and short side. Binance's own futures product didn't exist before September 2019, which covers more than half that study's sample window.

Why this one closed without a Binance test

Every other closed hypothesis in this series—Grid and Momentum—went through the project's own replication before reaching a verdict. This one didn't, and the reason is specific: when the longest, most recent, and most methodologically careful available study already answers the question of statistical significance in the exact regime that matters, running a private replication would just re-derive a result the literature has already established with more data than this project could practically collect on its own.

It's also a different kind of failure than the earlier two. Grid's problem was a fee tier: fixable, in principle, with the right account. Momentum's problem was costs eating a real but marginal spread: fixable, in principle, with lower fees or a bigger edge. Trend's problem, for the one universe actually accessible on Binance today, isn't fixable by any amount of capital or execution skill. The evidence that looks positive belongs to a period of market history that has already happened. No account size buys a way back into 2019.

Full write-up, with the complete regime-split numbers and references: Trend: When Published Edge Ages →. Also available as a working paper on SSRN 7405060 → and at DOI 10.2139/ssrn.7405060 → (preprint, not yet peer-reviewed).