University of Twente Student Theses


A machine learning approach to one-step overnight interest rate forecasting

Caspari, Christina (2021) A machine learning approach to one-step overnight interest rate forecasting.

[img] PDF
Abstract:Alternative Reference Rates (ARRs) are benchmark interest rates for overnight lending and include the Euro Short-Term Rate (€STR), Secured Overnight Financing Rate (SOFR) and Secured Overnight Index Average (SONIA). Ernst & Young has a variety of clients with financial contracts that are referenced with these rates. Regularly, clients request to receive a cash-in and cash-out flow estimate already in advance of the ending of the interest period. The required overnight rate is then not available yet as it is only published each business day. We calibrated different types of ARIMA and Random Forest models to test which model yields the highest forecasting accuracy of €STR, SOFR and SONIA. We developed one-step prediction models to lay a foundation for research in overnight interest rate forecasting. Furthermore, we included market and mathematical input features to increase the complexity of the models. We found that including input features to ARIMA decreases mean absolute error (mae) for €STR prediction from 19.9 bps to 5.48 bps, for SOFR from 68.8 bps to 34.6 bps and for SONIA from 5.99 bps to 4.91 bps. Including input features into Random Forest decreases the mae for €STR prediction from 240 bps to 16.9 bps and for SOFR from 67 bps to 9.25 bps. The performance of SONIA is less accurate using a Random Forest with input features as compared to without input features with an increase in mae from 6.22 bps to 35.5 bps. Furthermore, we found that the prediction accuracy of a Random Forest is heavily influenced by the choice of parameter settings. We decreased the mae by 12 bps for €STR, 4.1 bps for SOFR and 1.8 bps for SONIA by changing the parameter settings of the Random Forest. Finally, we have compared the performance of ARIMA and Random Forest to a simple prediction model with the assumption that the value of the interest rates of any given day is equal to the value of the interest rate of the previous day. All models that include input features produce error measures lower than the simple model.
Item Type:Essay (Master)
Ernst & Young GmbH Wirtschaftsprüfungsgesellschaft
Faculty:BMS: Behavioural, Management and Social Sciences
Subject:83 economics
Programme:Industrial Engineering and Management MSc (60029)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page