Many mathematical models used by investors to predict stock market movements were severely tested by this year’s collapse in share prices following the COVID-19 outbreak. But sophisticated techniques can now identify the key variables that influence a particular prediction and thus allow those factors to be overridden by humans.

During crises, users of models trained on more normal market conditions tend either to replace their carefully-designed investment process with ad hoc decisions made while emotions are running high, or they turn off the model and ignore the signals.

But there has been huge progress in the ability to explain the output of complex machine-learning models. Models can now be used to gain insight, rather than just make predictions – a huge change from the old-fashioned ‘black box’ models, where opaque relationships led to either anarchy or paralysis during a crisis.

HSBC’s model was trained using data that includes the extreme conditions of the 2008-09 financial crisis. However, most of the training data comes from more normal conditions, so some of the relationships learned will not perform well during crises.

The model thus under-predicted the likelihood of a shares sell-off early this year. Forecasting an exogenous shock like the coronavirus is immensely challenging for a model, but progress in model explainability means we can add a discretionary overlay that means models remain valuable during unprecedented market events.

In essence, we can tell which of the input variables were most responsible for a prediction and see whether they increased or decreased the forecast. Analysis can show which features the model typically finds most useful and thus how each influences the predictions – which make the model more, or less, bearish.

The three most important variables were: whether investment analysts were upgrading or downgrading companies, if US economic data was generally better or worse than expected over the last month, and changes in US bond markets.

In contrast, information about investment flows and positioning, though often viewed as being very informative of the market sentiment that drives prices, proved to be over-rated for predicting global equities.

So a model can now build up an internal score-card identifying which indicators are low-risk and which flag danger zones – then use it to make predictions.

We can see that profits revisions were already well into the danger zone before the COVID-19 outbreak disturbed markets. And while the model has learned that lower US bond yields are supportive for shares, when yields collapse because of a strong risk-off move they cannot provide that support.

US economic data was outperforming expectations – but during crises, the ‘expectation’ from official forecasts typically lags the true expectation of the market. Some forecasters are slow in adjusting their predictions and even when they do update, they often underestimate the true impact on the measured data.

Developments in model explainability thus permit us to construct a discretionary overlay that allows us to add human insight of how markets and how economic data are likely to respond as a crisis unfolds. By doing this we can still use our model as a framework for thinking about markets.

First published 18 March 2020.

Would you like to find out more? Click here to read the full report (you must be a subscriber to HSBC Global Research).

Disclosure and Disclaimer

More, collapsed
Join the conversation?

Join our Linkedin group to get an unparalleled view of macro and microeconomic events and trends from a bank that is a leader in both developed and emerging markets.