Background
A health crisis, with over 100 million confirmed infections and over one million deaths and counting. An economic crisis, with millions becoming unemployed and GDP falling by unprecedented magnitudes (the Office for Budget Responsibility estimates 2 million jobs lost and a 35% drop in GDP in the UK). And a social crisis, leading to increased loneliness and mental health issues, together with spikes in domestic violence, divorces and suicides. These are truly unprecedented times.
Under these extraordinary circumstances, the rapidly evolving conditions and government regulations make it even harder for organizations of all kinds to quickly adapt and avoid collapse. And now, when machine learning is needed the most, it fails to deliver.
The reason is that traditional machine learning (ML) works by learning patterns observed in past data. ML models can work very well when the new data that is fed in looks similar to the data they have been trained on. When modelling the real world, however, this is hardly ever the case, and this problem is especially pronounced during a crisis.
A paradigm shift in the science of ML is required so models will be able to promptly and efficiently adapt to changing conditions. Understanding causality is the key to achieving this goal. A machine that truly understands the relationship between cause and effect only requires a few data points to make an accurate assessment about a new regime, much in the same way a domain expert would. An additional benefit is that truly causal models are able to reduce cognitive biases and are even able to identify such biases present in the underlying data.
This is the philosophy that powers Causal AI, enabling it to forecast based on a much more profound understanding of the world, rather than on just statistical analysis. In this way, Causal AI is able to detect any sudden change of regime and alert its user. Subsequently, it autonomously rediscovers the causal structure of the new reality and uses this information to make predictions that reflect the emerging situation.
Case Study
To demonstrate how Causal AI works in practice, we present the following case study based on a financial dataset. This example illustrates how correctly identifying causal drivers not only results in improved performance but, more importantly, in faster model adaptation, especially in times of crisis.
Data
This case study uses hundreds of variables to predict Moody’s AAA Corporate Bond Yields one day ahead. We used a range of macroeconomic variables, including interest rates; stock market indicators, including returns and dividend yields; as well as other alternative datasets. As monetary policy in times of crisis drives interest rates even lower, high grade corporate bonds are one of the few low-risk vehicles to obtain a yield from cash.
Predicting bond yields can help investors navigate the markets, financial managers manage their firms’ exposure to interest rates, and economists to model the global economy.
Results with Current State of the Art
The latest automated machine learning (AutoML) techniques were used as a baseline. These include online learning, robust cross-validation, regularisation, model parsimony and meta-learning.
Forecasts proved to be reasonably accurate during stable times, with a mean absolute percentage error of 1.06% up to 24/02/2020. However, as the crisis hit the financial markets and volatility started to increase, these conventional AutoML models produced increasingly larger errors, proving them highly inadequate for modelling in the new regime.
Results with Causal AI
Forecasts with Causal AI also proved to be accurate in times of stability, achieving a similar level of accuracy to the current state of the art models in machine learning (in fact, with 7% higher directional accuracy). Nonetheless, where Causal AI really excelled was during volatile times. Causal AI was able to detect the change of regime and rapidly adapt to the new conditions in only a few data points, yielding great results when needed the most.
Causal AI achieved much lower error rates across various metrics, as well as a 18% higher directional accuracy than current machine learning during the crisis period. More importantly, it adapted significantly quicker, taking only 1/3 of the time to perform as accurately as during stable times, and find the causal relationships governing the new regime.
While current machine learning used a total of 10 features to predict bond yields, Causal AI settled on 3 main drivers causing the changes in bond yields. The greater simplicity of the Causal AI models offers advantages in model explainability and interpretability, as well as consuming lower computational resources and preventing overfitting
Causal AI models adapted 3 times quicker to new conditions than models built with current state of the art ML technology
The case study presents just the tip of the iceberg of what the science of causality and Causal AI have to offer. Causal AI isn’t just relevant for financial markets, but for every type of time-series dataset, including healthcare data.
Causal AI models produced 18% higher directional accuracy than models built with current state of the art ML technology
Summary of Findings
- Current state of the art machine learning can perform reasonably well during stable markets, however, it fails to perform during crises, when it is most needed.
- Causal AI identifies causal relationships that provide better forecasting accuracy and allow models to rapidly adapt to a constantly changing world.
- The ability of Causal AI to provide accurate predictions, even during periods of crisis, enable you to optimise your business and maximise your bottom line.
Download our White Paper
This paper sets out how Causal AI adapts 3 times quicker to crisis conditions than models built with current state of the art machine learning technology.
