Time-Varying Liquidity Reserve Forecasting with Layered ML
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Financial institutions face challenges in accurately predicting their liquidity needs due to evolving regulatory requirements and increased payment uncertainty, leading to potential insufficiencies in liquidity reserves that can cause significant risk and disruption.
Innovation Solution
A computer-implemented method using a predictive model with a first time-series layer and a second dense layer to process and forecast time-varying reserves, incorporating time-series variables and cyclic transformations to enhance transaction data, and utilizing dynamic scaling and feedback mechanisms for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional liquidity management methods are used, then operational simplicity is maintained, but prediction accuracy of liquidity needs deteriorates
Solution Approach 1:
The patent introduces time-series variable intermediaries (e.g., rolling averages, exponential smoothing components) that mediate between raw transaction data and liquidity predictions. These intermediaries capture temporal patterns without requiring complex manual analysis, thereby improving prediction accuracy while maintaining manageable system complexity.
Solution Approach 2:
The patent replaces traditional mechanical liquidity management approaches (manual forecasting, static buffers) with data-driven time-series analysis. This substitution enables dynamic, accurate predictions of liquidity needs based on historical transaction patterns, eliminating the need for overly simple rule-based systems while achieving high prediction accuracy.
2Reliability
If liquidity buffers are increased to ensure sufficiency, then reliability of debt settlement is improved, but operational efficiency and cost performance deteriorate
Solution Approach 1:
The patent implements dynamic liquidity buffer management where buffer sizes adjust automatically based on predicted transaction patterns and liquidity needs. Rather than maintaining static, oversized buffers, the system dynamically optimizes buffer levels to ensure debt settlement reliability while minimizing excess liquidity holdings, thereby improving operational efficiency.
Solution Approach 2:
The patent incorporates feedback mechanisms where actual liquidity consumption is compared against predictions, and buffer levels are adjusted accordingly. This closed-loop approach ensures reliable debt settlement by maintaining adequate buffers while preventing excessive liquidity accumulation, thus balancing reliability with operational efficiency.
3Measurement precision
If detailed transaction data is analyzed to improve prediction accuracy, then measurement precision is improved, but data processing time and computational resources increase
Solution Approach 1:
The patent extracts and focuses on the most relevant time-series features from transaction data (e.g., trend components, seasonal patterns, rolling statistics) rather than processing all raw data in detail. This selective extraction maintains high prediction accuracy by capturing essential patterns while significantly reducing data processing time and computational overhead.
Solution Approach 2:
The patent applies partial analysis by focusing computational resources on the most influential time-series variables and recent data points, rather than exhaustively analyzing the entire historical dataset. This approach achieves sufficient prediction accuracy for practical decision-making while minimizing processing time and resource consumption.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
There is described a computer implemented method for predicting a level of a time-varying reserve stored by a financial institution. The method comprises: obtaining a raw time-series transaction data set comprising a plurality of pieces of transaction data, each piece of transaction data corresponding to a transaction carried out by the financial institution over a previous time period; generating an enhanced time-series transaction data set by associating a plurality of time-series variables with each piece of transaction data in the raw transaction data set; providing the enhanced transaction data set to a predictive model trained using a machine learning method, wherein the predictive model comprises at least two layers, a first time-series layer configured to capture time-series nature of the enhanced transaction data set and a second dense layer configured to map outputs from the first time-series layer to predicted time-series data over a future time period; and receiving, from the predictive model as output, a prediction of the level of the time-varying reserve for the future time period. A computer implemented method for training the predictive model is also described.