Automated Data Processing of Bank Statements for Cash Balance Forecasting

Abstract

The forecasting of cash inflows and outflows across multiple business operations plays an important role in the financial health of medium and large enterprises. Historically, this function was assigned to specialized treasury departments who projected future cash flows within different business units by processing available information on the expected performance of each business unit (e.g. sales, expenditures). We present an alternative forecasting approach which uses historical cash balance data collected from standard bank statements to systematically predict the future cash positions across different bank accounts. Our main contribution is on addressing challenges in data extraction, curation, and pre-processing, from sources such as digital bank statements. In addition, we report on the initial experiments in using both conventional and machine learning approaches to forecast cash balances. We report forecasting results on both univariate and multivariate, equally-spaced cash balances pertaining to a small, representative subset of bank accounts.

Improving the Accuracy of Forecasting the TSA Daily Budgetary Fund Balance Based on Wavelet Packet Transforms

“…As shown in Figure 3, the wavelet-based forecasting procedure works best for large samples comprising more than 300 training points (time steps). A model was developed and considered in this research for forecasting cash balances in the 2019 federal budget TSA based on DWT using Daubechies family MW (DW) from DW(1) to DW (10) to achieve the set goal and solve the corresponding tasks. Our A model was developed and considered in this research for forecasting cash balances in the 2019 federal budget TSA based on DWT using Daubechies family MW (DW) from DW(1) to DW (10) to achieve the set goal and solve the corresponding tasks.…”

“…A model was developed and considered in this research for forecasting cash balances in the 2019 federal budget TSA based on DWT using Daubechies family MW (DW) from DW(1) to DW (10) to achieve the set goal and solve the corresponding tasks. Our A model was developed and considered in this research for forecasting cash balances in the 2019 federal budget TSA based on DWT using Daubechies family MW (DW) from DW(1) to DW (10) to achieve the set goal and solve the corresponding tasks. Our findings contribute to the achievement of the goal and confirm or refute the working hypotheses H1 and H2, assuming: (1) a DWT-based preliminary decomposition of the time series, compiled from the daily cash balance in the federal budget TSA, improves the accuracy level of traditional forecasting methods; (2) the choice of the parent discrete wavelet packet transform of the Daubechies family at the stage of time series decomposition will improve the forecasting accuracy from 80% to more than 96%.…”

“…Academic science has poorly studied the research area related to cash flow and balance forecasting, as evidenced by the results of recent scientific publications [10,11]. These publications show the usefulness of forecasting in cash management; however, they fail to consider improving the accuracy of the forecasts used in the respective models.…”

“…In particular, forecasting the dynamics of a nonlinear and non-stationary time series is associated with some significant volatilityrelated problems due to the influence of seasonal and calendar factors and the mutual influence of various other factors (macroeconomic, social, political, technogenic, financial, market, etc.). Reasons are the complexity, unpredictability, and multi-scale nature of the environmental impact, which did not find due attention in [10][11][12][13][14][15][16][17][18][19][20][21][22][23][24][25].…”

“…html (accessed on 1 May 2020))). For daily cash balances in the 2019 federal budget, TSA used DWT-based time series decomposition (from DW [1] to DW [10]) concerning different values of the decomposition level (from 1 up to 8) on a sample of 300 tested points with a forecast to a depth of 55time steps. As a result, they are characterized by a fairly high level of accuracy (more than 96%), which indicates the achievement of the research purpose and the validity of hypotheses (H1) and (H2): the DWT-based preliminary time series decomposition contributes to achieving a forecasting accuracy level of more than 96%.…”