Is a Recession Coming? What Economic and Financial Market Data Are Telling Us Respectively

With the Russia-Ukraine war triggering a commodity price shock and the Fed on a hawkish tightening path to tame inflation, fears about an impending recession have been on the rise recently. The brief inversion of the Treasury 2y10y curve, a widely watched recession signal, added further to the fears in early April. Are we getting close to a recession in the US or the Euro Area? How reliable is the 2y10y curve inversion in predicting an upcoming recession? To answer these questions, we looked into modelling recession probabilities over different time horizons using both economic data and financial market data and assessed our model performance through out-of-sample back-testing. According to our analysis, models using financial market data currently are indicating higher near-term recession probabilities than models based on economic fundamentals. However, our back-test results have also shown that historically models using financial market data were more likely to give false positive predictions for recessions than models using economic variables, especially over short time horizons. In other words, we think market fears for an imminent recession are likely overdone, though there is more uncertainty for longer horizons such as in 2-3 years.

Forecasting Recessions: Our Methodology

We constructed classification models using different predictors to predict whether there will be a recession in the next 6-month window, 1-year window, 2-year window or 3-year window in the US and the Euro Area respectively. Our models for each geography included separate single-variable models using financial market variables (yield curve steepness, equity performance, credit spread etc.), one composite model using all the financial market variables, and one composite model using purely economic data (PMIs, consumer confidence, M1 growth etc.). We split labelled historical data into multiple sets of training and testing data so that we could compare models’ out-of-sample performance. In terms of model selection, we used both the more traditional Logistic Regression model and the more advanced Explainable Boosting Machine (EBM) classifier model. Results from the better performing one based on out-of-sample performance for each model construction and each forecasting horizon are presented below (see more details on our methodology in the Appendix).

For the US we used data since July 1963, which covers 8 recessions as defined by the NBER. For the Euro Area we used data since December 1986, which covers 4 recessions as defined by the EABCN.

What Are the Latest Model-Implied Recession Probabilities and How Reliable Are the Models?

In our view a good recession-forecasting model needs to strike a balance between giving few false alarms and not missing actual recessions. Hence, we compared model performance across the following three metrics.

Accuracy: calculated as correct predictions/total predictions, which tells us how likely the model predictions are right.

Precision: calculated as true positive predictions/ (true positives + false positives), which tells us how likely an actual recession occurred when the model predicted a recession, in other words, whether the model gives a lot of false alarms.

Recall: calculated as true positives/ (true positives + false negatives), which tells us how many actual recessions the model could recall out of the total number of recessions, in other words, if the model tends to miss targets.

We summarise the latest predicted recession probabilities over 6M, 1Y, 2Y and 3Y using different models and the models’ past performance metrics in the tables below.

As shown in the first table, US economic variables are suggesting low probabilities of a recession in 6M or 1Y, though the 1Y probability has been rising slightly over recent months (see Chart 1). In contrast, most single-financial-variable models are suggesting above 30% probabilities of a near-term recession, much higher than implied by economic variables. However, based on our back-test results as recorded in the second table, models using single financial variables historically tended to have low precision, i.e., more likely to raise false alarms for near-term recessions. The composite model using financial variables had more reliable predictive performance in the past, and its implied near-term recession probabilities dropped from a brief peak of above 50% in mid-March to below 30% now – closer to the implied probabilities from the model using economic variables.

One drawback of the composite model using economic variables is that it had lower recall historically, i.e., it is more likely to miss true positives, especially for the 2Y and 3Y horizons. In other words, there is still high uncertainty regarding recession probabilities over the longer time horizons despite recent strong economic data.

US
Latest Recession Probabilities by Different Models

US
Model Out-Of-Sample Performance Metrics

Above, yellow selection: High near-term recession probabilities indicated by single-factor models using credit spread, VIX, SPX or Oil, but these models historically had lower precision, i.e. more likely to give false positives.

Above, blue selection: Composite model using economic variables is indicating lower recession probabilities across the forecasting horizons, likely due to still very strong data. However composite model using economic variables historically had lower recall than composite model using financial variables, i.e. more likely to miss true positives, especially for longer forecasting horizons (2Y, 3Y). In other words there is higher uncertainty over the 2-3 year horizon regarding whether a recession will happen than implied by latest economic data.

Probability of US Recession (using Composite Economic Variables)

Probability of US Recession (using Composite Financial Variables)

Similarly for the Euro Area the implied near-term recession probabilities from the composite model using economic variables are lower than models using financial variables. This is also after a temporary spike in recession probabilities from models using financial variables in March. In other words, financial markets moved quickly in March to price in a potential near-term recession, but since then have been re-assessing and shifted closer to probabilities implied by fundamental economic data. At the same time, economic variables are also moving to indicate rising recession probabilities for the 1Y and 2Y horizons. Nevertheless, one caveat here is that due to the much shorter history of data available the Euro Area’s models in general tended to have much less robust performance historically compared to the US ones.

Eurozone
Latest Recession Probabilities by Different Models

Eurozone
Model Out-of-Sample Performance Metrics

Above, blue selection: For the Eurozone, near-term recession probabilities indicated by models using financial variables spiked in March and have dropped lower recently to be more in line with probabilities implied by the composite model using economic variables. In general Eurozone models using financial variables tended to give more false positive predictions historically; especially for SX5E.

Probability of EU Recession (using Composite Economic Variables)

Probability of EU Recession (using Composite Financial Variables)

Conclusion

We constructed machine learning models to investigate recession probabilities implied by economic variables and financial variables respectively. Given still strong economic data, our models suggest low likelihood of a near-term recession for both the US and the Euro Area, though 1Y recession probabilities have been rising slightly over recent months. In contrast, financial markets moved quickly in March to price in high near-term recession probabilities, but recently reversed some of the pessimism. However, recession risk is more elevated over the longer time horizons of 2Y and 3Y, which recent economic data might not fully reflect.

Appendix: How We Constructed Our Recession Models

Data preparation

We first retrieved the recession dates for the US and Europe at the monthly frequency, which would form the basis of our recession labelling. Since we were interested in forecasting recessions at the 6-month (6M), 1-year (1Y), 2-year (2Y), and 3-year (3Y) horizons, we tagged the corresponding months before the starting month of each recession with a positive label. For example, if the recession occurred in 2000-07, and we were interested in the 6M horizon, we would use positive labels for each month starting from 2000-01 (6 months before 2000-07).

Feature engineering

After labelling our target variable of interest (“recession” or “no recession”), we retrieved 2 types of variables for both the US and Europe: financial and economic variables. The variables were grouped as such because we were interested to better understand if financial variables outputted different recession signals compared to economic variables. Financial variables used include yield curve steepness, equity performance, credit spread and oil prices. Economic variables used include PMIs, Consumer Confidence, Industrial Production, Monetary Aggregates and so on. A total of 100+ variables were used.

We used the Explainable Boosting Machine (EBM) to assist in feature selection, along with general economic/financial intuition, to create parsimonious models. Multiple time-series imputation methods were then used to fill in missing data, ranging from simple forward-filling to more complicated methods such as training time-series models to impute the next missing value. Finally, various transformations (e.g., differencing) were applied on the variables to make them more “stationary”.

Model training

An important characteristic of the data is that recessions tend to be quite rare – which is surely preferred in real-life, but unfortunately not so in a modelling context due to class imbalance (i.e., many non-recession labels and few recession labels). To mitigate this issue, we used Synthetic Minority Over-sampling Technique (SMOTE), including its more non-linear variants such as SVM-SMOTE, to balance the number of samples for each label.

To aid with model interpretability, we trained various “glassbox” models such as Logistic Regression (LR) and the Explainable Boosting Machine (EBM). In addition, we also trained multiple single-variable models to compare their performance against multi-variable models. Finally, each forecasting horizon also had its own set of models.

To ensure that the models were trained without lookahead bias (i.e., they were not able to “peek” into the future), we used an expanding window horizon to train each model and then make out-of-sample predictions on whether there would be a recession across the various forward-looking horizons (6M, 1Y, 2Y, 3Y). We also made sure that there was at least 1 recession in each train-test iteration, so that the model would be able to “learn” the patterns in the data that were associated with a recession.

We also ensured that the final set of months (e.g., last 6M till now for the 6M horizon) were removed from the training data as data labelling for those months depends on a future event (whether a recession happens next month) which we do not have absolute certainty on.

Model evaluation

Given the class imbalance issue, accuracy was not the most appropriate evaluation metric to use (e.g., if 90% of the samples were labelled as “no recession”, a naïve model that always predicted “no recession” would be 90% accurate). In addition, we were also interested in the precision and recall of the models to better understand if the models were giving false positive signals and/or missing actual recessions. As such, we used the F1-score, which neatly balances precision and recall, as our preferred evaluation metric.

Model inference

The best performing models based on F1-score were chosen for each of the forecasting horizons, across each geography, and for each variable type (i.e., whether the model used financial or economic variables). Using the latest trained models, new daily data for the financial variables was used as input for the financial variables model, while new economic data (at a lower frequency, e.g., monthly) was used for the economic variables model.

---

Tao Pan – Head of AI & Big Data

Lee He – Data Scientist

Jacob Pang – Data Scientist Intern

---