Deep Learning Applications in Multidimensional Cash Flow Forecasting

Abstract

Forecasting cash flows is a crucial task in corporate cash management. Accurate cash flow forecasts allow cash managers to (1) optimally allocate idle cash to short term investments and (2) minimize risk of insolvencya situation that can prove quite costly. Classical methods in cash flow forecasting make various assumptions of sta- tionarity, normality, and linearity that recent research has shown to be unreliable. Such methods also focus on single bank accounts where real-world use cases often involve several accountseach designated for specific transactions and counterparties. These designations create correlations between accounts and cause nonlinear patterns in data previously underutilized in forecasting. In this thesis, we outperform classical cash flow forecasting methods. We build on recent research into nonlinear modeling approaches by applying deep learning models and multi-account correlations to provide superior forecast accuracy. Our methods produce 5 day near-term forecasts with as much as 81% lower error than the naive estimator, where previous research achieved 33% lower error. Our results confirm the immense practical value in nonlinear approaches. Looking forward, we hope to pave the way for further applied research in the cash flow forecasting space. This will be possible through the future construction of a user interface and continued work with our confidential research partner.