Document Type : Original Article
Authors
1
Ph.D. in Strategic Management, University of Tabriz, Tabriz, Iran.
2
M.A., Tarbiat Modares University, Tehran, Iran.
3
PhD in Business Management, University of Tehran, Tehran, Iran; Department of Business Management, Nabi Akram Higher Education Institute, Tabriz, Iran
Abstract
Predicting the movement of financial resources into and out of corporate banking clients is a critical challenge for liquidity risk management and strategic treasury planning. Traditional forecasting models rely predominantly on linear statistical techniques and fail to capture the complex, dynamic interactions between firms and their banking relationships. This study introduces a novel predictive framework grounded in Neural Potential Field Networks (NPFNs). This hybrid architecture integrates artificial neural networks with the theoretical principles of potential field theory drawn from physics and robotics. Formally, each corporate client i at time t is represented as a point in a financial state space, and the model learns a scalar potential function whose negative gradient defines an attraction-repulsion vector field over that state space; a feedforward neural network parameterizes this potential function, and the gradient driving each client's predicted fund flow is obtained by automatic differentiation of the learned potential with respect to the client's state vector, conceptually analogous to artificial potential field methods in robotic motion planning. The proposed model conceptualizes corporate clients as dynamic agents operating within a financial potential field landscape, wherein fund inflows represent attraction forces and outflows represent repulsion forces acting upon liquidity reservoirs. By adapting the gradient-based mechanics of potential fields to model directional liquidity flows between enterprises, the framework captures both short-term transactional volatility and long-term structural patterns in corporate fund behavior. The model produces monthly fund flow forecasts at the individual client-account level over a 12-month out-of-sample evaluation window. The model is trained and validated on a panel dataset comprising 12 large Iranian commercial banks, covering the period 2018 to 2023 and encompassing over 3,400 corporate client accounts. Empirical results demonstrate that the NPFN model achieves a Mean Absolute Percentage Error (MAPE) of 4.73% for inflow prediction and 5.21% for outflow prediction, outperforming LSTM, GRU, and traditional ARIMA benchmarks by statistically significant margins, with all models evaluated on an identical held-out 12-month test period (January-December 2023), an identical feature set, and hyperparameters tuned via grid search on a validation subsample drawn from the preceding 12 months of the training window. Additionally, the model identifies key macroeconomic and microeconomic drivers, including interbank interest rate spreads, corporate leverage ratios, supply chain interconnectedness, and central bank regulatory signals, as primary determinants of potential field gradients. These findings provide actionable intelligence for bank asset-liability management (ALM) committees and offer a theoretically grounded, computationally tractable framework for next-generation corporate liquidity forecasting.
Keywords
