At the start of the year, the US President Donald Trump threatened to bomb Venezuela, Iran and Greenland. At the beginning of March, Trump, alongside Israel, initiated airstrikes against Iran, sending markets into a tailspin. The conflict, which started only with the US/Israel and Iran, has now spread to other countries, dominating headlines across the world. A constant flow of such news could send financial markets into a bear market and sometimes a crisis. This is why it is essential to invest in tools that help sift through this constant newsflow to detect where the financial environment has changed. This article unpacks our new, robust, real-time financial indicator.  

This financial indicator aims to directly address the shortfalls of existing indicator methods. For example, many indicators are only available months later, such as GDP-based indicators. Other indicators measure non-financial aspects of the economy, such as manufacturing output or interest rate changes. While useful, these proxies can be misaligned for price-specific events.   

The indicator’s building process 

This indicator had to meet a few criteria. Regimes, commonly referred to as bull and bear periods, needed to measure periods of similar global financial behaviour, often overlapping with known geopolitical events. For example, we might call Covid-19 a crisis or Brexit a bear environment. Regimes should last at least a few weeks but typically months. The model should at minimum be able to detect in real-time, otherwise known as nowcasting.  

We opted for a Gaussian Mixture Model (GMM), which is a machine learning technique. In short, this is a way to take unlabeled points and cluster them into sensible groups. In the picture below, we can take the random dark yellow points, tell the GMM that we think there are three clusters, and the model will identify the three groups on the right. The clusters it finds also matches are intuition for this 2D picture.  

Linking back to regime detection, GMMs can be used to identify economic regimes. Suppose the horizontal axis is return and the vertical axis is volatility. Each yellow dot would then be a specific week’s return/volatility. Based on the cluster’s position, we can assign labels. For example, the blue cluster has a high volatility and a high return, which we might describe as a rally.  

The diagrams shown so far are fictional data, but the same principles apply to our actual data. For each week, we can calculate statistics like return, volatility and more. The GMM then identifies clusters, and based on the clusters’ positions, we can assign labels. While humans might be able to do this in two dimensions, GMMs are able to cluster in five or many more dimensions.  

Data

This model used weekly returns from global equities going back to 1988. A considerable amount of data, which was made possible through our factor library ORCA, which was discussed in our last quant research piece titled Future Proofing Quantitative Investing for Complex Markets. From the weekly returns, we calculated 14 volatility features such as Sharpe ratios, drawdowns and autocorrelations. We then had several selection criteria aimed at finding the features capable of separating the data. This narrowed down to five features selected for the model: 

• Rolling volatility: The average volatility for companies in the previous month  
• Dispersion: Measures the spread of returns in the current week  
• Max drawdown: Average largest drawdown for companies in the last week 
• Explained variance: How well a single trendline explains recent returns 
• Change in 10-year US Bond Yield was added as an economic indicator 

Results

The model’s final predictions are shown above. At a value of two, we have the “crisis” cluster corresponding to many known crises such as 9/11, 2008 and Covid. At a value of one, we have the “bear” cluster, which is more frequent but less severe than a crisis. For example, the last two bears correspond to Trump’s November 2024 election and subsequent April tariffs. Lastly, along the zero-line, we have the “normal” cluster.

Based on the underlying features, we can describe the different clusters. Normal times are uneventful and low volatility; a safe environment to act on bets. However, when the market turns “bear”, there is increased uncertainty, volatility, and drawdowns. Returns are driven by other factors as opposed to their pure fundamentals. Lastly, we have crises. These are infrequent, but severe shocks are often characterised by extreme uncertainty, large losses, and erratic returns. The model does not find proof of “bull” markets, since the normal group is already limited and has strict volatility levels, so dividing it further does not enhance performance.

Algorithm enhancements 

GMM is not a time-series model and it had to be adjusted for a financial setup. A simple matching algorithm was used to match regimes predicted through time so that the regime called “one” in the 2000s is the same as the regime called “one” in 2025.  

The GMM does not understand that the predictions at each date are related to the weeks around it, which can cause the predictions to be quite jumpy. A hindsight smoothing algorithm is implemented to ensure sensible, sequential historical regimes. For example, the start and end dates of regimes are tidied. Short-lived regimes of less than six weeks are removed. Regimes with low probabilities are also removed. Removing historical noise helps us correctly classify future predictions.  

However, in real-time, we do not have hindsight, and so we need a decision rule for deciding how to interpret real-time predictions. For example, the plot below shows the probabilities of the different regimes over the first few months of 2025. Clearly, in real-time, views can be choppy, jumping from one week bear to normal and then to crisis. A decision rule allows us to consistently deal with these situations.  

During the back-test, careful attention was paid to limiting lookahead bias. At each date, the model was refitted using only the dates prior to that date, and its predictions were stored. We could then compare these “nowcast” predictions to our view a few weeks later. If the following weeks confirm the initial prediction, it is considered a true value, otherwise classified as a false alarm.  

Without the decision rule, the model has 315 switches predicted, corresponding to 110 false alarms, a hit rate of 65%. With the decision rule, this dramatically drops to 114 switches, of which only 10 are false alarms, corresponding to a hit rate of 91%. Even a simplistic decision rule of always waiting two weeks corresponds to a hit rate of 80%. 

Applications

A real-time regime detector has wide applications in business. Indicators can be used to manage tracking error budgets or sector constraints. We also investigated how different factor indices perform in different regimes. Below we have different MSCI Factor indices, as well as the RAFI Fundamental Index relative to the three regimes: normal, bear and crisis. The IR is calculated with respect to the MSCI ACWI benchmark. The results show Growth outperforming under Normal conditions, while Quality measures outperform in crisis – aligning with our intuition. Similar exercises can be done for asset classes, sectors, and more.

Conclusion 

The news is filled with noise, which is constantly filtering into the financial markets. A smoothed GMM was developed to filter through this noise to nowcast the current regime. Through various algorithms, we can find regimes which are long-lasting and sensible in real-time. Tools such as these open new opportunities in risk management, dynamic factor models and strategic asset allocation – resulting in more robust and situationally aware funds.