Anna Soloveva - Predicting the Net Charge-Off Rate of Bank Holding Company

Predicting the Net Charge-Off Rate of Bank Holding Company

The Goal of the Project

The goal of this project is to build a linear prediction model to estimate the Net Charge-Off Rate (NCO) from the commercial and industrial loan portfolio of Huntington Bancshares Incorporated under the basic and severely adverse macroeconomic scenario.

Huntington Bancshares Incorporated is a Fortune 500 company headquartered in Columbus, Ohio. The bank primarily operates across a seven-state banking franchise for Ohio, Illinois, Indiana, Kentucky, Michigan, Pennsylvania, and West Virginia. The commercial bank is a $109 billion asset regional bank holding company as of December 31, 2019.

Macroeconomic Variables and Transformation

To predict the Net Charge-Off Rate for Huntington Bank, various macroeconomic and financial factors were used including Real Gross Domestic Product (RGDP), Real Disposable Income Growth (RDIG), Unemployment Rate (UR), Consumer Price Index (CPI), 5 Year Treasury Yield (T5YR), BBB Corporate Yield (BBBY), Mortgage Rate (MR), Prime Rate (PRIME), Dow Jones Total Stock Market Index (DJIND), House Price Index (HPI), Commercial Real Estate Price Index (CREIND) and Market Volatility (VIX). CPI is used to estimate inflation in the economy. Please refer to Table 1 to observe the associated performance of these factors with Net Charge-Off Rate.

In addition to these factors, Spread and TSpread variables were generated. Spread is the difference between BBBY and 10 Year Treasury Yield. Variable TSpread represents the difference between 10 Year and 3 Month Treasury Yields.

Quarterly variables were created for Date, DJIND, CREIND, VIX, UR, and RGDP using the lag function in the R program. Annual unemployment rate was calculated using quarterly rates. Year-over-year (YOY) method of evaluation was applied on DJIND and RGDP to better understand the impact of stock prices on the economy.

The Huntington bank data ranges from the years of 1999 to 2019. It was collected from the Federal Financial Institution Examination Council (FFIEC). The historical macroeconomic data ranges from the years of 1976 to 2019, and was collected from the Board of Governors of the Federal Reserve System.

Model Selection Process

Three linear models were built to predict Net Charge-Off Rate for Huntington Bank. Different economic and financial factors (annual, quarter lags and year-over-year) were used to construct different models to better understand the relationship between the target variable and predictors.

A correlation matrix was conducted to choose the best predictors for the target variable. Table 2 lists the highest and the lowest correlated variables to the Net Charge-Off Rate. Different combinations of the highly associated factors were run together to build three predictive models. These models were then compared based on their respective 10-Fold Cross Validation (CV) score and Mean Squared Error (MSE).

Estimates for Linear Model 1

The Annual Unemployment Rate Quarter Lag 1, Commercial Real Estate Price Index, and Market Volatility Index Quarter Lag 1 were found to be the best explanatory variables to predict Net Charge-Off Rate in Model 1. All predictors were significant at 90% confidence level. The model’s adjusted R-squared value is 64.09%. In other words, the model explains 65.47% of variation in our target variable assuming everything else remains constant (Figure 1).

Interpretation of Coefficients

Commercial Real Estate Price Index: For every one unit increase in the index, the target variable decreases by 0.0009.
Annual Unemployment Rate Quarter Lag 1: For every one unit increase in the rate, the target variable increases by 0.1654.
Market Volatility Index Quarter Lag 1: For every one unit increase in the index, the target variable increases by 0.0075.

Estimates for Linear Model 2

The Nominal GDP YOY Quarter Lag 2, DJ RET YOY Quarter Lag 2, and DJ Stock Index Quarter Lag 1 were found to be the best explanatory variables to estimate Net Charge-Off Rate in Model 2. All predictors were significant at 90% confidence level. The model’s adjusted R-squared is 57.29% which suggests our variables explain 58.93% of variation in our target variable assuming everything else remains constant (Figure 2).
Interpretation of Coefficients
Nominal GDP YOY Quarter Lag 2: For every one unit increase in the index, there is a 1.578e-02 decrease in Net Charge-Off Rate.

DJ RET YOY Quarter Lag 2: For every one unit increase in the index, there is a 3.429e-03 decrease in Net Charge-Off Rate.
DJ Stock Index Quarter Lag 1: For every one unit increase in the index, there is a 2.192e-05 decrease in Net Charge-Off Rate.

Estimates for Linear Model 3

The Annual Unemployment Rate Quarter Lag 2 and Market Volatility Index were found to be the best explanatory variables to estimate Net Charge-Off Rate in Model 3. All predictors are significant at 90% confidence level. The model’s adjusted R-square value is 65.20% and the selected variables explain 66.09% of variation in our target variable assuming everything else remains constant (Figure 3).
Interpretation of Coefficients
- Annual Unemployment Rate Quarter Lag 2: For every one unit increase in the explanatory variable, there is a 0.1761 unit increase in the target variable.
- Market Volatility Index Quarter Lag 4: For every one unit increase in the explanatory variable there is a 0.0092 unit increase in the target variable.

Model Comparison

Displayed in Table 3, we can observe that Model 3 has the lowest Mean Squared Error and 10-Fold Cross Validation values. Therefore, it is considered to be the champion model and most superior of the three in predicting the target variable, Net Charge-Off Rate. In addition, charts were designed for each of the models under baseline and severely adverse economic scenarios . The charts represent the model's behavior under each of these scenarios along with its actual performance. Model 1 and 3 have a steep increase in the NCO under baseline scenario, while Model 2 predicts a decrease in the NCO. Under a severely adverse scenario, we can observe that Model 1 and Model 3 predict that NCO might rise to the level of the recession in 2008, however Model 2 prediction is much lower in magnitude.

Conclusion

The overarching goals were to create the best model for predicting the target variable, Net Charge-Off Rate. Historical and macroeconomic series data was used to conduct analysis and predict the impact of the target variable. Statistical methodology, economic theory and financial applications were applied to assessing data from the commercial and industrial loan portfolio of Huntington Bancshares Incorporated. Summary statistics were computed, along with plots to visually distinguish association. The correlation matrix aided in determining the best predictors and transformation of variables were generated as needed. As a result, Model 3 outperformed Models 1 and 2, due to its lowest Mean Squared Error and 10-Fold Cross Validation. For future research, an analyst can dive deeper to look at the impact on Net Charge-Off Rate that is influenced by a future risk, such as a recession or financial crisis.

Authors: Mia Fernandes, Tunisha Kansra, Jacqueline Pisano, Anna Soloveva