CUSUM is a sequential way to flag a persistent change in a series, not a crystal ball for the next market move. It adds small deviations from a chosen baseline until the accumulated evidence crosses a threshold; that alert depends on the baseline, threshold and data you feed it.
What CUSUM detects
A CUSUM (cumulative sum) procedure tracks whether observations keep leaning away from a reference level. A change point is a time at which the statistical behaviour of a series shifts; the procedure can raise an alarm after enough evidence accumulates, but it does not identify the cause or guarantee that the new behaviour will persist.
In a financial-surveillance study, Polunchenko and colleagues treated structural breaks as sequential anomaly detection and compared a Shiryaev–Roberts procedure with a CUSUM chart on real financial data. They report that both performed well in their case study, with their proposed procedure appearing slightly better (arXiv paper). That is a result for their setup, not a universal ranking of detectors.
The moving parts behind an alert
A useful CUSUM design makes three choices explicit: what is monitored, what counts as the reference behaviour, and how much accumulated deviation triggers an alarm. A one-sided version can watch for persistent moves above a reference; a separate one can watch below it. Changing the reference, scaling or threshold changes what the detector notices and how often it stays quiet.
observations → compare with reference → accumulate deviations → threshold crossed? → review alertFor trading research, the observed series might be returns, residuals from a model, or a volatility measure. These are not interchangeable: an alert in returns is not automatically an alert in volatility. If you monitor several assets together, a panel method can use cross-sectional information. A double-CUSUM study describes ordering CUSUM statistics across panel members and evaluates change detection in simulations, with an application to S&P 100 constituent returns (University of Bristol record). That is a different task from a single-series trading signal.
What the evidence does—and does not—show
The cited studies establish that change-point procedures can be formulated and evaluated on financial series; neither demonstrates that a CUSUM alert, used as a trading rule, produces net profits after execution costs. Detection quality and trading value are separate questions. A prompt alert may be noisy, while a cautious threshold may react only after much of a shift has passed.
There is also a model-risk problem. If the reference distribution is a poor fit, ordinary variation can look like a break—or a real change can be missed. Repeatedly adjusting the threshold after inspecting the same history turns the detector into another tuned strategy. The alert itself is statistical evidence conditional on modelling choices, not a label supplied by the market.
How you would test a CUSUM filter
Write down the monitored series, baseline, scaling, threshold and reset rule before looking at the final test period. Then compare the detector with a simple no-filter baseline and a clearly specified alternative. Measure both detection behaviour (false alarms and delay against a defined change label) and, separately, any trading outcome after spread, commission, slippage and the intended position rules.
Keep every tuning decision inside the training window. Freeze the rule before testing later data, and repeat across distinct periods rather than selecting the one chart where the alarm looks timely. The same discipline used to challenge backtest overfitting applies here. realbacktesting's proof process is built around results that can be inspected and reproduced; a detector deserves the same scrutiny, not a more forgiving one.
A backtest still cannot prove that a detected break has a stable economic cause. Regime labels are definitions chosen by the researcher, and simulated fills may not match live execution. Report the choices, missed alerts and false alarms alongside any strategy result.
Takeaway
CUSUM can make a persistent deviation easier to notice. It cannot tell you, by itself, whether that deviation is a durable regime or a profitable opportunity.