Building Trading Strategy Ensembles from Classified Market States

Summary

The document describes constructing strategies by classifying market states instead of directly forecasting prices or returns. Labels define states, and market features such as moving averages are selected for their ability to distinguish those labels. A separate neural network predicts each label from the features and outputs buy or sell probabilities; the recommendations are combined linearly into candidate strategies.

To select strategies for ensembles, the authors create a score based on historical returns that also penalizes strategies with few trades or little capital deployed. They report that ensembles built from the top-scoring strategies outperform individual comparisons in both returns and risk-adjusted returns out of sample on Bitcoin. They also find that the custom historical score correlates with future success. The document notes that some candidate strategies can be misleading, but gives no sample dates, benchmark details, cost assumptions, or numerical results. Because labels are calculated using future data, the method depends on keeping that information out of the classifier's inputs and evaluation process; the summary does not describe those safeguards.

Key ideas

• The method classifies labeled market states rather than directly forecasting returns.
• Features are chosen for their ability to separate the state labels.
• Neural-network classifiers produce buy or sell probabilities that are combined into strategies.
• A custom historical score penalizes low trading activity and limited capital involvement.
• Bitcoin ensembles reportedly beat the cited comparisons out of sample, though implementation details are not provided.

Tags

# Constructing trading strategy ensembles by classifying market states


# Constructing trading strategy ensembles by classifying market states









Rather than directly predicting future prices or returns, we follow a more recent trend in asset management and classify the state of a market based on labels. We use numerous standard labels and even construct our own ones. The labels rely on future data to be calculated, and can be used a target for training a market state classifier using an appropriate set of market features, e.g. moving averages. The construction of those features relies on their label separation power. Only a set of reasonable distinct features can approximate the labels. For each label we use a specific neural network to classify the state using the market features from our feature space. Each classifier gives a probability to buy or to sell and combining all their recommendations (here only done in a linear way) results in what we call a trading strategy. There are many such strategies and some of them are somewhat dubious and misleading. We construct our own metric based on past returns but penalising for a low number of transactions or small capital involvement. Only top score-performance-wise trading strategies end up in final ensembles. Using the Bitcoin market we show that the strategy ensembles outperform both in returns and risk-adjusted returns in the out-of-sample period. Even more so we demonstrate that there is a clear correlation between the success achieved in the past (if measured in our custom metric) and the future.

Shown in full with attribution under the source's licence. Licence: abstract CC0

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.