TL;DR
This paper introduces a novel "Boundless Model" (BM) for cash management that simplifies existing approaches by eliminating the need for predefined bounds, leading to guaranteed optimal solutions and significantly faster computation times. The BM is shown to be robust to forecasting errors, offering substantial cost-risk savings for companies.
What Problem Does This Paper Address?
Managing cash effectively is a critical task for businesses, involving daily decisions on how much cash to hold versus how much to invest. Traditional cash management models often rely on setting upper and lower bounds for cash balances. When a balance hits these bounds, a transaction (buying or selling investments) is triggered to bring it back within limits. However, these bounds introduce non-linearity into the problem, making it difficult to guarantee truly optimal policies. Existing methods, like evolutionary algorithms, can find good solutions but don't guarantee optimality. This paper aims to address this by developing a new model that can find optimal cash management policies without relying on these problematic bounds, while also being computationally efficient and robust to real-world forecasting inaccuracies.
Methods
The authors propose a two-pronged approach:
- Linearization and Reformulation: They first linearize the cash management problem by reformulating the transfer and holding/shortage cost functions using auxiliary binary variables. This allows for a more straightforward mathematical programming approach. They then apply this to reformulate the existing Gormley-Meade (GM) model into a linear-quadratic multiobjective program (MOGM). This reformulation, unlike the original GM model, guarantees optimality.
- The Boundless Model (BM): The core contribution is the "Boundless Model" (BM). This model takes the insights from linearization and further simplifies the problem by removing the need for explicit bounds altogether. Instead, it directly optimizes the policy (the sequence of transactions) to minimize a combined cost and risk objective. The BM is formulated as a quadratic programming problem, which can be efficiently solved to guarantee optimality.
Datasets
The study utilizes:
- 54 real cash flow data sets: These datasets are from small and medium-sized companies in Spain with variable lengths (170 to 1508 observations) and cover a period of about 8 years. This data is used for sensitivity analysis of the BM to forecasting errors.
Key Insights and Ideas
- Bounds are a bottleneck: The paper argues that the reliance on predefined bounds in traditional cash management models is a primary cause of suboptimal solutions and computational complexity.
- Linearization unlocks optimality: By linearizing the cost functions, the authors pave the way for guaranteed optimal solutions through mathematical programming.
- Direct policy optimization is superior: The "Boundless Model" demonstrates that directly optimizing the transaction policy, rather than optimizing bounds that determine the policy, leads to a simpler and more effective approach.
- Forecasting errors are manageable: The BM shows remarkable robustness to prediction errors, suggesting that companies can still achieve significant cost-risk savings even with imperfect forecasts.
- Computational efficiency: The BM offers a substantial speedup in computation time compared to both the reformulated GM model (MOGM) and the original GM model using genetic algorithms.
Key Results
- Optimality Guaranteed: Both the MOGM and the BM formulations guarantee optimal solutions, unlike the original GM model which uses approximate methods.
- Computational Speedup: The BM is significantly faster than the GM model using genetic algorithms. For example, in one experiment, the BM's run time was negligible (0.005 sec) compared to the GM model's run time (up to 969 sec for 10,000 generations).
- BM vs. MOGM Speed: The BM also shows faster run times than the MOGM, with average run times being 38% lower and standard deviations 41% smaller.
- Robustness to Forecasting Errors: A sensitivity analysis on 54 real datasets showed that the BM achieves substantial cost-risk savings even with significant forecasting errors (up to 40% of standard deviation). The savings are still notable even with near-random cash flows.
- Equivalence Theorem: The paper proves that the BM and the MOGM are equivalent in terms of the optimal solutions they produce, given the same inputs.
Limitations and Future Work
- Fixed Policy Assumption: The sensitivity analysis assumed a fixed policy for the entire planning horizon. In practice, cash managers would likely re-optimize the policy as new information becomes available. Future work could explore dynamic re-optimization.
- Multiple Bank Accounts: The current model primarily considers a two-asset setting. Real-world companies often manage multiple bank accounts with varying characteristics. Extending the BM to handle multi-account scenarios is a promising future direction.
- Robust Optimization: To further enhance immunity to unreliable forecasts, the authors suggest exploring robust optimization techniques.
- Complex Cost Functions: Incorporating more complex cost functions, such as time-dependent costs or piecewise functions, could lead to a more accurate representation of real-world cash management problems.
Summary updated