Structural Break Tests and Meta-Labeling for Strategy Regimes
Summary
The document addresses how traders might identify market conditions associated with changes in algorithmic strategy performance. It proposes structural-break analysis as one avenue, naming CUSUM tests on recursive residuals or levels and explosiveness tests such as Chow-type Dickey-Fuller and supremum augmented Dickey-Fuller methods. The answer offers these as tools for investigating changing behavior in market data; it does not provide an implementation, compare their detection performance, or show results for the trader’s strategies.
It also suggests adding a secondary machine-learning model to filter signals from a primary strategy. In meta-labeling, this second model estimates whether a primary signal is likely to be correct, and its confidence can inform position size. A trend-following example explains that sideways conditions may generate unprofitable crossover signals and transaction costs, while features such as volatility or autocorrelation could help identify those conditions. This is a proposed workflow rather than evidence that the filter improves returns. Regime tests and secondary models require careful validation to avoid mistaking noise for persistent conditions or fitting the same performance history they are meant to explain.
Key ideas
- Structural-break methods can help investigate whether strategy performance changes alongside market conditions.
- The answer lists CUSUM and explosiveness tests as possible tools for detecting breaks.
- Meta-labeling uses a secondary model to assess primary strategy signals and can feed confidence into position sizing.
- A trend-following example links sideways markets with false crossover signals and transaction costs.
- The proposed methods need validation because the document provides no measured evidence of improved performance.
Tags
Full text
# Extract market features to decide when to deploy or stop strategies
# Extract market features to decide when to deploy or stop strategies
I have been live trading using algorithmic strategies for a year. I have good periods, lasting about two months, followed but bad periods of few weeks. I did the necessary statistical tests to ensure that good periods are not the result of sheer luck, but rather consistent and significant performance.
I am now wondering whether I could identify market regimes that are correlated with my trading performances. This would provide a tool to decide when to start/stop my algos. I will start with what seems obvious to me: volatility and asset correlations.
I have little experience in this field so I would like to find standard methods if they exist, papers or discussions.
## Answer by Jacques Joubert (score 0, accepted)
https://quant.stackexchange.com/a/45325
I think you are right in saying that under certain market states your models perform poorly. There are a few useful tools that you could make use of.
Starting off we look at useful financial features, one being structural breaks.
Structural breaks are generally split into two groups (de Prado 2018):
- CUSUM Tests
- Explosiveness Tests
For CUSUM tests you can have a look at:
- Brown-Durbin-Evans CUSUM test on Recursive Residuals.
- Chu-Stinchcombe-White CUSUM Test on Levels, which follows Homm and Breitung 2012
For Explosiveness Tests:
- Chow-Type Dickey-Fuller Tests, Chow, 1960
- Supremum Augmented Dickey-Fuller, Peter C. B. Phillips Yangru Wu Jun Yu, 2011
- Sub- and Super-Martingale Tests, Greene, 2008
It would be even better if you train a secondary ml model to help filter out the false positives. I have advocated for this technique before:
"Whilst reading this I realized that it would be a really good application for meta-labeling. The idea behind meta-labeling is to build a second model that determines if the signals {0, 1} from the primary model are correct or not.
By doing this the secondary model outputs a value between 0 and 1 indicating how confident the model is that the primary model is correct or not. This output can then be passed to a bet sizing algorithm which maps the output to a position size. The core idea being that we want to take large positions on trades that are likely to be true and smaller positions on trades when we are unsure.
To give some intuition behind this. Lets take a trend following strategy as an example. Now moving average crossover strategies are known to under perform when the market moves sideways. The choppy nature causes a lot of transaction fees.
The secondary model will pick up that under some volatility conditions and perhaps a low auto correlation, that we are in a side ways trend and thus the primary models signal (a 1 in this case) is likely to be false and so it assigns it a low probability.
More about this technique can be read about in Chapter 3 of Advances in Financial Machine Learning. A toy example of meta-labeling can be found here "Meta-Labeling on MNIST Data."Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.