Trend following

The idea

Momentum — the tendency for assets that have risen to continue rising, and assets that have fallen to continue falling — is one of the most replicated findings in financial markets. It exists across asset classes, across time horizons, and across geographies.

Trend following is the systematic implementation of this observation. A trend system measures the direction and strength of recent price movement and takes a position in that direction. It holds the trade while the trend persists and exits when momentum reverses.

What the evidence says

The modern benchmark. Moskowitz, Ooi and Pedersen (2012) tested 58 liquid futures and forwards across equity indices, bonds, currencies and commodities, and found time-series momentum returns positive in all of them and statistically significant in most. A market's past 12-month return predicted its next month's return. The effect faded after about a year and then partly reversed, a pattern the authors say fits slow reaction to news followed by late overreaction. They found no link to measures of investor sentiment, and trend returns in different asset classes moved together more closely than the asset classes themselves.

More than a century of data. Hurst, Ooi and Pedersen (2017) rebuilt the test back to 1880 across 67 markets, hand-transcribing early commodity prices from Chicago Board of Trade annual reports. The pattern held through the Great Depression, two world wars and 1970s stagflation. They link trends to anchoring and herding, and to central banks and corporate hedgers who trade for reasons other than profit. They also trace the idea to David Ricardo's advice to cut losses short and let profits run, recorded in 1838.

Who sits on the other side. CFTC position data in the 2012 study show speculators riding trends for about a year and cutting back as trends turned, with hedgers on the opposite side. The authors conclude that speculators appear to profit from time-series momentum at the expense of hedgers, who use futures to manage business risk.

The sceptics. Huang, Li, Wang and Zhou (2020) found little evidence of time-series momentum when they tested market by market, and judged the large pooled test statistic unreliable. They also found the strategy's results virtually the same as those of a similar rule based on each market's historical average return, which needs no predictability. Kim, Tse and Wald (2016) found that much of the measured edge came from volatility scaling, which sizes each position to the same risk. Without it, the strategy's cumulative results looked much like buy-and-hold.

Time-series versus cross-sectional

Time-series momentum compares an asset's current price to its own history. If BTC is above its 200-day moving average, the time-series signal is long. The trade is about whether the asset is trending relative to itself.

Cross-sectional momentum compares assets to each other. The system ranks all assets by recent performance and goes long the top performers, short the bottom — regardless of whether any individual asset is in an absolute uptrend.

Both signals are used. Time-series momentum drives directional conviction. Cross-sectional momentum drives relative allocation between assets in the same universe.

The smile and its limits

The smile. Plotted against S&P 500 quarters, the 2012 study's trend returns form a smile, with the best results in the biggest up and down quarters. Fung and Hsieh (2001) had already modelled this option-like shape. In their model a trend follower behaves like the holder of a lookback straddle, an option that pays the distance between the highest and lowest price over a period.

Which markets did the work. In their sample, which ended in 1997, Fung and Hsieh found that trend funds' returns in the most extreme equity months were explained by trends in currencies, short-term rates, bonds and commodities. Lookback straddles on stock indices did not help explain them. In the industry data they cite, stock indices drew the least trend-following activity of any market. The authors warn that the attribution was made after the fact, so its statistical inference is not reliable.

Slow bears, fast crashes. In their simulation, Hurst, Ooi and Pedersen found trend positive in most of the ten deepest drawdowns of a 60/40 portfolio since 1880. Those drawdowns took about 15 months on average from peak to trough, which gave trend systems time to turn short after the first leg. The authors name the October 1987 crash as a fall too fast for the strategy to turn short in time.

The mechanism. Dao et al. (2016) showed that a trend position earns roughly the gap between realised variance over the trend's own horizon and realised variance over its rebalancing interval. A steady fall builds a large move from moderate daily changes, and the trend earns. A shock that lands inside one rebalancing interval counts on both sides of that gap, so the trend is flat on average while a straddle bought beforehand pays. The authors conclude that a six-to-nine-month trend system rebalanced weekly should not be expected to hedge a crash lasting a few days.

Long volatility, in part. Fung and Hsieh described trend followers as performing as if they were long volatility. Hurst, Ooi and Pedersen found results similar whether market volatility was rising or falling. Dao et al.'s split helps reconcile the two, since a trend book is long variance over its own horizon and short variance over its rebalancing interval. A straddle buyer pays a fixed premium up front, while a trend follower pays as it goes, through the small whipsaw losses of ranging markets.

Why regime matters

Trend strategies work in trending regimes and lose in ranging regimes. When a market chops sideways, the system repeatedly enters a trend that immediately reverses, accumulating small losses across many trades.

Regime detection manages this. When signals are inconsistent — short-term and long-term momentum pointing in opposite directions, volatility elevated, breadth compressed — position sizes reduce. The system runs small until the signals align again.

Choosing the signal speed

Goulding, Harvey and Mazzoleni (2023) frame signal speed as a trade between two errors. A slow signal, such as 12 months, filters noise but reacts late at a turning point. A fast signal, such as one month, reacts quickly but often fires on noise. In the working-paper version, covering the US stock market from 1969 to 2018, the two disagreed in more than a third of months, so mixed signals are routine. The authors' answer is to blend slow and fast signals, which partly offset each other when they disagree.

Speed also changes the shape of returns. Hamill (2016) at Man Group found positive skew strongest for faster signals, and Hurst, Ooi and Pedersen found that a one-month delay hurt the short-term signals most. Taken together, the signals with the most positive skew are also the most exposed to delay. Lempérière et al. (2014) found the trend effect stable across two centuries of data, with no sign of decay in trends of several months in the recent period. Much shorter trends, of about three days, had weakened markedly since 1990.

The filter itself matters less. Levine and Pedersen (2016) showed that time-series momentum and moving-average crossovers are equivalent in their most general forms, and a crossover is essentially a fixed blend of time-series momentum signals. They suggest that managers focus on implementation, such as costs, position sizing, diversification and risk management, more than on which filter to start from. A 200-day average and a 12-month return are close relatives, which may help explain why so many managers hold similar positions.

Where it breaks

Sharp reversals. A long trend can end in a single session that wipes out weeks of accumulated gains. Stop-loss levels determine how much profit is surrendered before the position exits. This give-back is the structural cost of the strategy.

Crowding. The most common trend signals — 200-day moving averages, 12-month momentum — are held simultaneously by many systematic managers. When they all exit at the same time, the reversal amplifies.

Low-volatility ranging markets. In calm sideways conditions, trend signals generate many small false starts. Individual losses are small but accumulate. Regime-based sizing limits the damage in these periods, but cannot eliminate the cost entirely.

What is different in crypto

Shorter horizons. Liu and Tsyvinski (2021) found strong time-series momentum in bitcoin at daily and weekly frequencies, with one week's return predicting the next one to four weeks. In the working-paper version, their bitcoin data run from 2011 to May 2018, and the evidence for ether, over a shorter sample, was weaker and mixed. The classic futures evidence sits at one to 12 months. A 200-day average on bitcoin runs far slower than the horizons where this study found the effect.

Longs pay to hold the trend. Schmeling, Schrimpf and Todorov (BIS, 2023) found the futures premium over spot in bitcoin and ether averaged above 10% a year and reached up to 60%, much larger than carry in equities, bonds, currencies or commodities. They tie it to trend-chasing smaller investors seeking leveraged upside, and to scarce arbitrage capital. A long held through dated futures gives up that premium as the contract converges on spot, and a long perpetual swap pays funding whenever the rate is positive. High carry also predicted price crashes in their data.

No halts, thin books. Crypto-native venues trade around the clock without the automatic halts used in other markets. On 10 October 2025, a threatened 100% tariff on Chinese imports set off more than $19bn of liquidations in roughly a day, and bitcoin's top-of-book depth on key venues shrank by more than 90%. A trend signal built on daily closes cannot respond to a move that starts and ends within hours.

Winning positions can be closed for you. When liquidations exhaust an insurance fund, most perpetual venues fall back on auto-deleveraging, which closes or cuts profitable positions on the other side to cover the shortfall. Chitra (2025) traces the usual ranking rule, profit multiplied by leverage, to BitMEX in 2016 and reports that venues with well over 95% of perpetual volume use it, though some venues use other rules. Chitra also reports that Hyperliquid used auto-deleveraging repeatedly on 10 October 2025, in a 12-minute cascade involving about $2.1bn of positions. A trend follower who is short into a crash with high leverage and a large open profit sits near the front of that queue.

The daily close is a choice. A market that never shuts has no official close, so every study and system has to pick a cut-off time. Liu and Tsyvinski measure each day's return at a fixed clock time, while the CME CF Bitcoin Reference Rate, which settles CME bitcoin futures, is set from an hour of trading that ends at 4pm London time. CF Benchmarks also publishes a New York variant set at 4pm New York time. Two systems running the same 200-day rule on different cut-off times can sit on opposite sides of the line when price is close to it.

Traps that are easy to miss

The years after a crisis. Moskowitz, Ooi and Pedersen show trend losing sharply from March to May 2009, as the crisis ended. Hutchinson and O'Brien (2014), using almost a century of data, found trend returns weak for about four years on average after a global crisis began, as the serial correlation of futures returns fell. In regional crises, results began to recover in the third year. The authors wrote after strong 2008 results drew new institutional allocations to trend followers, whose returns then ran below their long-term average.

Two readings of the 2010s. Hurst, Ooi and Pedersen linked weaker trend results after 2008 in part to high correlation across markets, which lasted from late 2008 to mid-2014. When markets move together, a volatility-targeted book cuts every position and holds fewer separate bets. A later AQR study, Babu et al. (2020), ruled out a lack of diversification as the cause of the weak 2010s and traced it mainly to fewer large market moves. It found no decline in the strategy's ability to turn trends into profits.

Cross-sectional crashes come from the short side. Daniel and Moskowitz (2016) found that equity momentum's worst losses cluster in panic states and arrive as the market rebounds. In July and August 1932 the US market rose 82% and past losers rose far faster, a pattern that returned from March to May 2009, the same months in which time-series trend lost sharply. After a crash, the loser portfolio's market beta can exceed 3, so in bear markets the strategy behaves like a written call option on the market. A crypto relative-strength book that is short the worst-hit coins after a sell-off can carry a similar exposure.

Crowding leaves traces. Lou and Polk (2022) gauge crowding in stock momentum through the excess correlation among the stocks the trade would hold, which they call comomentum. When it is low, momentum trading tends to correct underreaction, and when it is high, the returns on momentum stocks later strongly revert. The BIS saw signs that commodity trading advisers and other momentum traders, selling after losses on their cross-asset positions, helped amplify the February 2018 sell-off. The S&P 500 closed more than 10% below its 26 January close by 8 February, and the BIS also suggested that CTA de-risking may have exacerbated intraday swings in oil and other commodities.