The question
Technical analysis has a poor reputation among economists, yet the evidence that simple trend-following rules forecast the stock market is surprisingly strong. Neely, Rapach, Tu and Zhou (2014) showed that moving-average, momentum and volume-based signals forecast the U.S. equity risk premium at least as well as the macroeconomic variables usually studied. A second robust finding is that return predictability is concentrated in bad times, usually measured by NBER recessions.
Recessions are an awkward conditioning variable, though. They cover only about a tenth of the post-1960 sample, and they are dated by a committee after the event, so an investor cannot use them in real time. We ask a related but more usable question: does the forecasting power of technical indicators depend on the level of financial uncertainty?
Calm and turbulent regimes
We measure uncertainty with the financial uncertainty index of Ludvigson, Ma and Ng (2021), which captures how unpredictable a broad panel of 147 financial series is. Its two largest spikes are Black Monday in October 1987 and the 2008 financial crisis. It is only moderately correlated with other uncertainty and sentiment measures; the highest correlation, with macroeconomic uncertainty, is 0.68.
We split the sample into low- and high-uncertainty regimes in two ways. The simple rule labels a month as high-uncertainty when the index is above its sample mean, which puts 43% of months in the high regime. The second approach fits a two-state Markov-switching model to the monthly change in the index. A likelihood-ratio test cannot reject equal means across the two states, so we use the version in which only the variance switches, and classify a month as high-uncertainty when its smoothed probability exceeds one half; this puts 46% of months in the high regime. Both definitions capture most NBER recession months (87% and 74%), but they are far broader than recessions, which make up only 12% of the sample. That leaves an investor many more months in which the distinction is useful.
What we do
Data. We use monthly S&P 500 prices and trading volume from 1960:08 to 2018:12, with returns measured in excess of the one-month Treasury bill rate. From these we build the 14 technical indicators of Neely et al. (2014): six moving-average rules, which signal “buy” when a short moving average of prices (one to three months) is above a long one (nine or twelve months); two momentum rules, which signal “buy” when the price is above its level nine or twelve months earlier; and six on-balance-volume rules, which apply the moving-average logic to volume-weighted price changes.
In sample. Each indicator enters a univariate predictive regression for next month’s excess return, with wild-bootstrap inference. To compare regimes we compute “disentangled” R² statistics, which measure how much of the return variation the regression explains within the low- and high-uncertainty months separately, as Neely et al. (2014) do for recessions and expansions.
Out of sample. Forecasts are produced recursively with an expanding window that starts with 181 months, matching Neely et al. (2014). Each forecast is compared with the historical mean using the Campbell–Thompson out-of-sample R², over the whole period and within each regime, with Clark–West tests.
Volatility. We also forecast monthly realised volatility, built from daily returns, using an exponentially weighted moving average of the predictive regressions’ residuals. Accuracy is measured by the proportion of the variability in realised volatility that the forecasts explain, with block-bootstrap p-values.
What we find
Finding 1. Over the full sample, technical indicators look unremarkable.
In sample, most indicators are significant at the 10% level, with R² values between 0.16% and 0.96%. Out of sample they are mediocre: several have positive R², but only one, the MA(2,12) rule (0.80%), beats the historical mean at the 5% level.
Finding 2. In calm markets they work; in turbulent markets they do not.
Split by regime, the picture changes. Using the Markov-switching regimes, every one of the 14 indicators explains more of the return variation in low-uncertainty months than in high-uncertainty months; for MA(2,12) the in-sample R² is 2.69% against 0.24%. Out of sample, all 14 have a positive R² in low-uncertainty periods, 13 of them significant at the 5% level or better, with values as high as 3.35% for the VOL(2,9) rule. In high-uncertainty periods most are negative, and the few that are positive are never significant. The simpler above-the-mean split is noisier, especially for the volume rules, but tells the same story for the moving-average and momentum rules.
- Low uncertainty
- High uncertainty
Notes: Regimes from the Markov-switching (variance) model. Every low-uncertainty value shown is significant at the 5% level or better (Clark–West); no high-uncertainty value is significant. Six of the 14 indicators shown. Source: Table 6.
Data table
| Low uncertainty | High uncertainty | |
|---|---|---|
| MA(1,12) | 2.22 | −0.47 |
| MA(2,12) | 2.32 | 0.02 |
| MA(3,9) | 2.47 | −1.37 |
| MOM(9) | 0.88 | −0.27 |
| VOL(2,9) | 3.35 | −1.55 |
| VOL(3,12) | 2.15 | −0.26 |
Finding 3. The same pattern holds for volatility.
Technical indicators also forecast volatility better in calm periods. Under the Markov-switching regimes, the forecasts explain about 49–51% of the variability in realised volatility in low-uncertainty months, significant at the 1% level for every indicator, against about 36–39% in high-uncertainty months, where none is significant. Over the full sample (about 39–41%), none of the volatility forecasts is significant either.
| Indicator | Full sample | Low uncertainty | High uncertainty |
|---|---|---|---|
| MA(1,9) | 39.98 | 48.98*** | 37.42 |
| MA(2,12) | 38.88 | 49.19*** | 35.95 |
| MOM(12) | 40.52 | 49.05*** | 38.09 |
| VOL(3,12) | 39.56 | 50.54*** | 36.43 |
Proportion of explained variability (Blair, Poon and Taylor, 2001); regimes from the Markov-switching (variance) model. *** denotes significance at 1% (stationary block bootstrap). Four of the 14 indicators shown. Source: Table 7 of the paper.
Why uncertainty matters
Why would trend-following signals break down when uncertainty rises? Cujean and Hasler (2017) show that disagreement among investors spikes in bad times and that this polarisation of opinion is what generates predictability then. A trend-chasing signal relies on many investors acting on similar information in the same direction, so it is natural for it to fail when opinions diverge. Two related mechanisms point the same way. When uncertainty is high, the most uncertainty-averse investors reduce or stop their participation in the stock market (Bali, Brown and Tang, 2017), and investors move into safer assets, a flight to quality that interrupts the price trends the indicators rely on.
The result does not contradict the evidence that predictability is concentrated in recessions. The effect is specific to technical indicators: we find no such difference for the macroeconomic variables of Welch and Goyal (2008). Recessions are also when the benchmark itself does worst. Squared forecast errors from the historical mean are 144% larger in recessions than in expansions, so it is easier to beat there. Robustness checks using a median split, or more than two regimes, give the same conclusion.
Takeaways
- Technical indicators forecast the equity premium in calm markets. Their predictive power is concentrated in low-uncertainty periods.
- They should be avoided when uncertainty is high. In turbulent periods their forecasts are no better, and often worse, than the historical mean.
- Uncertainty is a usable, real-time state variable. Unlike recession dates, it can be measured as events unfold, and it covers far more months.
- The practical rule is simple. Check the uncertainty regime before acting on a technical signal.
Key references
- Bali, T. G., Brown, S. J., and Tang, Y. (2017). Is economic uncertainty priced in the cross-section of stock returns? Journal of Financial Economics, 126(3), 471–489.
- Blair, B. J., Poon, S.-H., and Taylor, S. J. (2001). Forecasting S&P 100 volatility: The incremental information content of implied volatilities and high-frequency index returns. Journal of Econometrics, 105(1), 5–26.
- Campbell, J. Y., and Thompson, S. B. (2008). Predicting excess stock returns out of sample: Can anything beat the historical average? Review of Financial Studies, 21(4), 1509–1531.
- Cujean, J., and Hasler, M. (2017). Why does return predictability concentrate in bad times? Journal of Finance, 72(6), 2717–2758.
- Hamilton, J. D. (1989). A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica, 57(2), 357–384.
- Ludvigson, S. C., Ma, S., and Ng, S. (2021). Uncertainty and business cycles: Exogenous impulse or endogenous response? American Economic Journal: Macroeconomics, 13(4), 369–410.
- Neely, C. J., Rapach, D. E., Tu, J., and Zhou, G. (2014). Forecasting the equity risk premium: The role of technical indicators. Management Science, 60(7), 1772–1791.
- Welch, I., and Goyal, A. (2008). A comprehensive look at the empirical performance of equity premium prediction. Review of Financial Studies, 21(4), 1455–1508.