Trend Following and Momentum Trading: A Complete Guide

Trend following (time-series momentum trading) is the single most documented anomaly in systematic finance: assets with positive trailing returns tend to deliver positive forward returns over horizons of one to twelve months. This is not a chart pattern — it is a measurable, slowly-decaying risk premium with a convex payoff and well-understood failure modes. This guide is about how to estimate it, size it, and avoid the ways it quietly stops working.

The canonical academic result (Moskowitz, Ooi, Pedersen 2012; Hurst, Ooi, Pedersen on a century of data) is that a simple sign-of-past-return rule produces a Sharpe near 0.7–1.0 gross across a diversified futures book. The catch every practitioner learns: that number is before turnover costs, before crowding, and before the multi-year drawdowns that have dominated the post-2010 sample. The edge is real but thin, and most of the work is in not destroying it.

What is actually being estimated

Strip away the indicators and a trend signal is an estimate of the conditional drift of returns given recent realized returns. The two workhorse specifications:

• Time-series momentum (TSMOM): signal is sign(r_{t-k:t}) for each asset

independently. You are betting an asset's own trailing return predicts its own forward return.

• Cross-sectional momentum (XSMOM): rank a universe by trailing return, long

the top decile, short the bottom. You are betting on dispersion between assets, which is closer to statistical arbitrage and carries strong negative skew ("momentum crashes," Daniel & Moskowitz).

These are not the same factor. TSMOM has long volatility / "crisis alpha" behavior; XSMOM has crash risk on sharp reversals. Most managed-futures programs are predominantly TSMOM because its convexity is what allocators pay for.

A moving-average crossover, a Donchian breakout, and a sign-of-return rule are monotonic transforms of the same underlying state — they are ~90% correlated once you control for lookback. Choosing among them is second-order; choosing the lookback, the volatility estimator, and the rebalance frequency is first-order.

Signal construction that survives contact with data

The right way to combine lookbacks is not to pick one. Use a panel of horizons and average the standardized signals:

import numpy as np
import pandas as pd

def tsmom_signal(close: pd.Series, lookbacks=(21, 63, 126, 252)):
    """Volatility-normalized, multi-horizon time-series momentum signal in [-1, 1]."""
    ret = np.log(close).diff()
    sigs = []
    for lb in lookbacks:
        mom = np.log(close).diff(lb)                      # trailing log return
        vol = ret.rolling(lb).std() * np.sqrt(lb)         # scale to same horizon
        sigs.append((mom / vol).clip(-2, 2))              # standardized, capped
    raw = pd.concat(sigs, axis=1).mean(axis=1)
    return np.tanh(raw).shift(1)                          # squash + lag to avoid look-ahead

Two non-obvious choices matter here. First, standardizing by horizon volatility turns the raw return into something comparable across assets and regimes — without it, your signal is dominated by whatever is most volatile. Second, the tanh squash caps position size on extreme readings; raw z-scores let a single asset dominate the book exactly when it is most likely to reverse.

Design choice	Common default	Why it matters	Failure if wrong
Lookback	single 12M	sets which cycle you harvest	overfit to one regime
Vol estimator	rolling std	normalizes risk	std lags vol spikes (use EWMA)
Rebalance freq	daily	controls turnover	daily = cost bleed on slow signal
Signal cap	none	limits concentration	one name blows up the book
Universe size	<10 markets	drives diversification	tiny N = regime coin-flip

Volatility targeting is most of the Sharpe

The largest single improvement to a naive trend system is ex-ante volatility targeting — scaling each position so its forecast contribution to portfolio volatility is roughly constant. This is not a tweak; in TSMOM research the volatility-scaled version has a materially higher and more stable Sharpe because it deleverages into turbulence and re-levers in calm.

def vol_targeted_position(close, signal, target_vol=0.15, halflife=33, cap=2.0):
    """Scale a [-1,1] signal to a target annualized vol using EWMA realized vol."""
    ret = np.log(close).diff()
    ewma_vol = ret.ewm(halflife=halflife).std() * np.sqrt(252)
    lever = (target_vol / ewma_vol).clip(upper=cap)       # cap leverage in calm regimes
    return (signal * lever).shift(1).fillna(0)            # lag: trade on next bar

Use EWMA, not a simple rolling window, for the volatility forecast: a trailing 20-day std responds to a vol spike with a step function and a cliff 20 days later, injecting artificial turnover. The leverage cap is essential — in ultra-low-vol regimes target/realized explodes and a naive scaler will put on ruinous size right before a vol regime change.

Honest performance: where the edge isn't

Trend following has had a brutal 2011–2019 stretch (the "CTA winter") where diversified programs were flat-to-down while equities rallied. This is not noise to be explained away — it is the structural cost of a strategy that needs sustained directional moves and gets chopped up by mean-reverting, central-bank-anchored markets. You will hold this strategy through multi-year drawdowns or you will not capture its convexity.

Where it has little or no edge after honest accounting:

• Single equities. Idiosyncratic single-name "trend" is mostly noise plus

short-horizon reversal; cross-sectional equity momentum is a real factor but is heavily crowded and decays fast.

• Short lookbacks on liquid futures. Anything under ~1 month is dominated by

microstructure and turnover costs; the gross signal is real but net is negative for most retail cost structures.

• High-turnover daily rebalancing of a slow signal. You pay spread and impact

to express a view that changes monthly. Trade the signal at the frequency the signal actually changes.

Signal decay is measurable: regress forward returns on the lagged signal and track the information coefficient (IC) over rolling windows. A trend IC that was 0.04 in the 1990s and is 0.01 today is telling you the premium is being arbitraged, not that your code is broken. Indicators that retail communities treat as edge — fast MA crosses, RSI, MACD — are simply low-IC, high-turnover features; test them with the same IC/decay lens and most fail to clear costs.

Diversification is the only free lunch

A single-market trend system is a bet on one regime; its Sharpe is dominated by sampling luck. The CTA model works because trend signals across ~50–100 weakly-correlated instruments (rates, FX, commodities, equity indices, crypto) diversify the idiosyncratic timing error while preserving the common crisis-alpha exposure. The realistic ceiling: gross Sharpe scales roughly with sqrt(Neffective) where Neffective is far smaller than your instrument count because asset-class blocks co-move in risk-off.

Crypto specifics

Crypto trends are violent and the signal IC has historically been higher than in mature futures, but three frictions dominate:

• Funding on perps is a real carry cost on multi-week holds; net it against

the trend PnL or you will overstate returns. See carry trade.

• Vol-of-vol is extreme, so EWMA vol targeting matters more here than

anywhere — see measuring volatility and volatility targeting.

• Capacity is small and crowding is fast. Crypto trend edges decay in

quarters, not decades; assume the signal you read about publicly is already half-arbitraged.

Capacity and crowding

Trend following is capacity-constrained at the slow, liquid end and crowding-constrained at the fast end. When too much capital chases the same breakout, the move that used to follow-through instead reverses as everyone's stops cluster — turning the convex payoff concave. Monitor your own market impact: if your fills are consistently worse than your backtest's mid-price assumption, you are the liquidity, not the taker.

Validation that won't lie to you

The 50/200 that looks perfect in-sample is the classic overfit. Defenses that actually work:

• Walk-forward, not single split. Re-estimate parameters on a rolling window

and trade only out-of-sample — see walk-forward optimization.

• Deflate for the search. If you tried 40 lookback/asset combinations, your

best Sharpe is inflated; apply deflated Sharpe and multiple-testing corrections.

• Model costs first, signal second. Put realistic spread, impact, and funding

into the backtest before you admire any equity curve, and watch for backtesting biases.

Judge results on risk-adjusted terms — Sharpe, Sortino, and especially Calmar / max drawdown, since trend equity curves have long, deep dips that a Sharpe alone hides.

Sizing the convex payoff

The payoff is positively skewed: many small losses, rare large gains. This breaks naive Kelly sizing, which assumes a stable return distribution — practitioners trade a small fraction of Kelly because full Kelly on a fat-tailed, autocorrelated return stream produces ruinous variance and path dependence. The discipline that matters most:

1. Take every signal. The edge lives in the rare large winner; cherry-picking

"obvious" trades deletes exactly the trades that pay for the year.

1. Size by ex-ante volatility, never conviction.
2. Trail, don't target. A fixed profit target truncates the right tail that

is the entire source of edge — see position sizing.

Conclusion

Trend following is a thin, convex, slowly-decaying premium — not a chart trick. The real work is volatility-scaled sizing, broad diversification, honest cost modeling, and the institutional stamina to sit through multi-year drawdowns. It is the structural opposite of mean reversion and pairs naturally with breakout exits. Treat published lookbacks as starting hypotheses to be re-estimated out-of-sample, assume your edge is decaying, and you will keep what little of the premium survives crowding and costs.