Predicting Capital Structure Using Machine Learning Techniques: Evidence From Sri Lankan Listed Companies

Abstract

"Over the past few decades, there has been much discussion on how the debt and equity is mixed in financing the business in the corporate finance literature. However, there isn't currently agreement on a trustworthy estimating technique for target leverage. Previous studies have demonstrated nonlinear relationships between leverage and factors determine the capital structure. However, as far as the Sri Lankan context is considered, there were not any research conducted in corporate finance domain for predicting the capital structure using machine learning techniques. To fill that gap and in order to find complicated patterns in the data, this project use machine learning approach. Results of the machine learning models outcome are then compared with conventional method (OLS) to see if they improve prediction performance.

The data set used to test this comprise of 40 companies distributed among 11 industries listed in Colombo stock exchange (CSE) for the specific period from 2010 till 2019. The financial data which were gathered comprise of debt to book value of total asset (TDA) (considered as the leverage) as the dependent variable and independent variables that are firm specific such as year’s profitability, firm size in terms of assets, maturity, market-to-book value, asset growth, tangibility of assets, cash, depreciation, corporate tax rate, and risk factor (Z- score), industry specific variable namely industry leverage, industry leverage, and macro variables namely, inflation and GDP.

This project employed algorithms of machine learning namely Support Vector Regression (SVR), Least absolute shrinkage and selection operator (LASSO), Artificial Neural network (ANN), Random Forests (RF), and Gradient Boosting Regression Trees (GBRT) to compare the performance with the results of the traditional approach (i.e - ordinary least squares model (OLS)). Previous research has demonstrated that machine learning properties are especially beneficial in enhancing predictive accuracy when complex patterns in training data become visible. For comparing the predictive performances of both OLS and machine learning techniques, code based predictive models were developed and prediction was performed using both models. Accordingly, the predictive performance was evaluated based on RMSE and R2.

It was identified The Random Forest outperformed the results of all of other machine learning techniques and superior to the OLS techniques with RMSE to 0.0178 (RF) from 0,0225 (OLS) and R2 to 0.4092 (RF) from 0,2508 (OLS). It was also noted that the R2 and RMSE generated from Gradient Boosting Regression Trees (GBRT)) Model is also more similar to the results of RF model. However, the result of LASSO model is less performing compared to the OLS method even though it is one machine learning model. It was also noted that according to the Random Forest machine learning techniques, the Z–Score, the Market to Book Value, Depreciation, Firm Size, and Tangibility as the most important determinants. And the project result suggest that Capital structure cannot be predicted using a single, straightforward model; instead, a combination of several factors is thought to be crucial for predicting the leverage.

In overall, this project resulted that the machine learning approach is able to improve the performance on predicting the capital structure (leverage). Furthermore, it was also noted that despite of the contribution made to the Sri Lanka corporate finance literatures, applying more machine learning techniques to predict the capital structure is suggested as a future work."