Introduction
A market regime is a market condition when market participants and transactions show specific behaviors and characteristics, e.g., risk aversion in a recession period. A simple (but not comprehensive) introduction can be seen in Ang et al. (2012). Regime study is important for the whole pipeline of quantitative trading, from raw alpha research to monetization, for the simple reason that it captures the fundamental trait of the market. Consider a mid-frequency strategy exposed more on growth stocks. It is natural to believe that a recession could cause a lower-than-usual alpha, a stylized alpha model, a more volatile portfolio, etc. This could also be an interesting source of alpha as we incorporate a general belief in the market.
Latent Modeling
One can easily categorize regimes by volatility, monetary policies, and other discretionary metrics. But more than that, regime modeling can be seen as discovering latent regime factors from market data from a (relatively) objective point of view. We can consider latent models such as a Hidden Markov Model (HMM) or a Gaussian Mixture Model (GMM). Botte et al. (2021) explores the use of GMM and shows a clear and interpretable result. Its idea is to first determine the number of latent regimes, then use GMM to determine the regimes as a mixture of Gaussian distribution from 17 risk factors. We can infer that other techniques that explain a latent pattern, such as clustering and VAE, can also be used. In general, the nature of regimes is latency and low dimensionality. There is a clear trade-off between subtlety and interpretability, and the latter is way more important in most cases.
Time Scales
As we go into a micro market research rather than a mid-to-low frequency research, we should be aware of the compatability between the time scale of regimes and the frequency of strategies. Typical open studies only use daily time series and provide a broad overview of markets. In HFT research, however, alpha models can be trained and updated using the most recent data in which regime changes can be too slow to be even included. It makes not much sense considering an annual regime change that happens for a strategy that is only trained with the past 1-month data, although that could also be a negligible feature for data selection.
Stylized Facts
THe HFT market schemes we’ve discussed by definition overlap with traditional market microstructures and day trading methology to some extent. For example, very often a micro market scheme comes with time. For equity markets, daily transactions concentrate after opening auctions and before closing auctions. For crypto and FX markets, different trading sessions are the most straightfoward schemes. Micro volatility is also a typical scheme pattern. Market schemes can be improved by better modeling (e.g., latent modeling) or microstructure study (e.g., traded volume description).
Prior Guidance
As we have pointed out from the beginning, market schemes can be used as a prior knowledge to guide research. A scheme such as a footprint chart feature can interact with simple feautures such as book pressure to significantly improve their qualities. A meta learning method can be applied to modeling, e.g., to train different models for different market schemes. Monetization, especially for market making strategies, is of course obviously related to market schemes.
Leave a Reply