Factor Timing

Once you accept Factor Investing — that value, momentum, quality and the rest each earn a premium — the natural next question is greedy: can you own each factor more when its payoff is about to be good and less when it's about to be bad? That is factor timing. Instead of a fixed 25% in each of four factors, you tilt the weights based on a signal: a valuation spread, recent factor momentum, a macro regime, sentiment.

The appeal is obvious and the trap is subtle. Factor premia are already noisy; timing them adds a second, even noisier layer of prediction on top. The academic consensus, most bluntly stated by Asness, is that factor timing is tempting and mostly a bad idea — the gains are small, fragile, and usually eaten by turnover and estimation error.

How timing signals are built

A timing signal maps some observable into a recommended factor weight. The most-studied is the valuation spread: how cheap the factor's long leg is relative to its short leg, measured by the difference in their valuation multiples.

$$\text{spread}_t = \text{value multiple of short leg} - \text{value multiple of long leg}.$$

The bet is that when a factor's long leg is unusually cheap versus its short leg (a wide spread), its future return is higher — a Mean Reversion view of the factor itself. Other timers use factor momentum (a factor that has done well recently keeps doing well for a while) or a regime classifier that switches tilts by macro state.

A weak but real predictor

The honest picture: even the best timing signals are weakly predictive. A valuation spread might genuinely forecast the next year's factor return, but with a low correlation — a small edge buried in a lot of noise. The scatter below lets you feel that. Drag the correlation up and down and watch how much a "real" relationship can still look like a cloud: at the modest correlations typical of factor-timing signals, the fit line is almost flat and the r² is tiny.

That low r² is the whole story. A signal can be statistically real and still explain almost none of next period's return — which means acting on it aggressively bets a lot of turnover on a faint edge.

Worked example: does timing beat holding?

You run a value tilt. Historically value earns 4% a year long-short with a volatility of 12% (a Sharpe of 0.33). You build a spread-based timer: when the value spread is in its widest quartile you double your value exposure, and when it's in its narrowest quartile you halve it.

Backtested, timing lifts the average return to 4.6% and cuts a little volatility, raising the Sharpe to about 0.40. Sounds like a win — until you count what it cost:

• Turnover. Doubling and halving exposure churns the book. If the extra trading costs 0.4% a year, your 0.6% of gross timing gain shrinks to 0.2% net.
• Estimation error. That 0.6% edge was measured on the same history you designed the timer on. Out of sample, spread-based timers have delivered a fraction of their backtest promise, and some none at all.

After costs and honest out-of-sample haircuts, the timed portfolio is barely distinguishable from just holding the factor — the standard result.

The one timer worth respecting, slowly

Not all timing is hopeless — it's fast, aggressive timing that fails. The valuation spread does carry information, but it works on a horizon of years and reverts slowly, so trading it hard just generates turnover. Used gently — a mild lean toward a factor when its spread is genuinely extreme (say a once-a-decade cheapness), rebalanced patiently — it can help. This also connects to Factor Crowding: a very compressed spread is a sign capital has piled in and the premium is thin, which is a real reason to trim.

The deeper lesson generalises beyond factors. Any time you stack a timing prediction on top of an already-thin edge, you multiply two noisy things and inherit both errors. Diversifying across factors and holding them patiently has repeatedly beaten trying to outguess which one is about to shine — a humbling but well-replicated result, and a close relative of the Alpha Decay and overfitting problems that haunt the whole field.