=Paper=
{{Paper
|id=Vol-3900/Paper17
|storemode=property
|title=Harnessing AI and Machine Learning for Improved Cash Flow Forecasting
|pdfUrl=https://ceur-ws.org/Vol-3900/Paper17.pdf
|volume=Vol-3900
|authors=Soumit Roy,Sayantan Polley,Saurav De,Chetan Gangwal,Chittresh Mitra
|dblpUrl=https://dblp.org/rec/conf/dosier/RoyPDGM24
}}
==Harnessing AI and Machine Learning for Improved Cash Flow Forecasting==
https://ceur-ws.org/Vol-3900/Paper17.pdf
Harnessing AI and Machine Learning for Improved Cash
Flow Forecasting
Soumit Roy1,† , Sayantan Polley2,† , Sourav De3,† , Chetan Gangwal4,† and Chittresh Mitra5,†
1
Jade Global Inc, Aurora, USA
2
Otto-von-Guericke-University Magdeburg, Germany
3
Government College of Engineering and Textile Technology, Serampore, Hooghly, India
4
Jade Global Inc, San Jose, USA
5
Dublin High School, Dublin, USA
Abstract
Accurate cash flow estimates are essential for organizations to maintain financial stability and optimize resource
utilization. When the economy undergoes fluctuations, conventional forecasting methodologies, which necessitate
human intervention and rely on historical data, are rendered ineffective. This investigation incorporates artificial
intelligence (AI) and machine learning (ML) to enhance cash flow forecasting. Utilizing robust algorithms and
real-time data processing, the technology generates precise predictions. Budgetary constraints may necessitate
modifications. The study addresses the following topics: system integration, model selection, engineering
characteristics, and data collection. The findings are corroborated by a variety of case studies from a wide range
of industries, which demonstrate significant improvements in forecast accuracy. Enhanced user interfaces and
visualizations may facilitate the decision-making process for stakeholders. AI enhances operational efficiency
and strategy development by assisting businesses in estimating cash flow and managing their finances.
Keywords
Cash Flow Forecasting, Artificial Intelligence, Model Selection, Machine Learning
1. Introduction
Cash flow forecasting is an essential component of financial management that is essential for
the maintenance of liquid assets, the assurance of timely payments, and the support of strategic
decisions [1]. Conventional methods that depend on historical data and human adjustments to
forecast future cash flows are exceedingly challenging, if not impossible, to implement due to the
complexity and volatility of contemporary financial markets [2]. Traditional methods may result in
imprecise projections that endanger a company’s financial stability [3]. These methods are potentially
cumbersome, inaccurate, and slow to adjust to evolving market conditions. Innovations in artificial
intelligence (AI) and machine learning (ML) have transformed numerous industries by introducing
cutting-edge solutions that enhance adaptability, efficiency, and accuracy. Financial forecasters have an
exceptional opportunity to leverage AI and ML to generate cash flow predictions that are significantly
more precise and dependable [4]. These technologies have the potential to enhance the precision
and accuracy of forecasts by processing large volumes of data in real time, recognizing intricate
patterns, and adapting to new data. Time series data processing is essential for job prediction, and
AI and ML models such as Transformer networks, GRUs, and LSTM networks have demonstrated
exceptional potential in this domain [5]. These models are more effective than traditional statistical
methods in capturing the long-term connections and patterns in financial data. The estimates are more
comprehensive and accurate because of AI and ML’s capacity to incorporate data from a variety of
The 2024 Sixth Doctoral Symposium on Intelligence Enabled Research (DoSIER 2024), November 28–29, 2024, Jalpaiguri, India
*
Corresponding author.
†
These authors contributed equally.
$ soumit.roy@ieee.org (S. Roy); sayantan.polley@ovgu.de (S. Polley); dr.sourav.de@gmail.com (S. De);
chetan.Gangwal@jadeglobal.com (C. Gangwal); chittresh15512@gmail.com (C. Mitra)
0000-0002-5532-346X (S. Roy); 0000-0002-0758-3021 (S. Polley); 0000-0001-8587-4062 (S. De); 0009-0008-6065-4969
(C. Gangwal); 0009-0001-7701-9098 (C. Mitra)
© 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
sources, such as external economic indicators, industry trends, and company-relevant features [5]. The
incorporation of ML and AI into cash flow forecasting is a multi-stage process that involves numerous
critical steps [6]. The initial stage should be to collect and organize consistent, high-quality input data.
The objective of feature engineering is to enhance the predictive capabilities of the model by identifying
and constructing pertinent components that significantly influence cash flow. A component of the
training and selection process is ensuring that the models have a strong performance, training them
using the most effective methods, and fine-tuning their hyperparameters [7]. Ultimately, stakeholders
can rapidly comprehend the outcomes because of user-friendly interfaces and visualization tools.
Real-time forecasting and seamless data transfer are enabled by the models’ integration with business
processes [8]. The reliability and accuracy of predictions have been significantly enhanced by the
proposed AI and ML-based approach to cash flow forecasting, which has been validated in numerous
case studies across a variety of industries. For instance, retailers have benefited from improved
inventory management and financial planning because of more accurate cash flow forecasts. Firms are
now capable of paying their suppliers promptly and optimizing their production processes because of
advancements in forecasting [9]. The service sector has experienced a decrease in waste and a more
efficient utilization of resources because of more precise cash flow projections. By integrating AI and
ML, businesses may be able to improve their cash flow predictions and obtain control of their financial
situation. This approach provides an advantage by facilitating the process of making informed de-
cisions, enhancing the stability of your finances, and accelerating operational efficiency and growth [10].
2. Literature Review
Khudhur et al.[11] proposed a method that founded on a meta-analysis of ten studies that investigate the
reliability of financial performance predictions derived by machine learning techniques. The analysis
focuses on issues and an examination of data quality and harmony. Random sampling and feature
extraction and selection are two methods for addressing data imbalance and high dimensionality,
respectively. The paper concentrates on the enhancement of model accuracy and robustness in the
prediction of financial stability and insolvency through data pretreatment, dimensionality reduction,
and balanced datasets. Malika Becha et al. [12] constructed a model that can categorize potential debtors
as having either outstanding or poor credit is the proposed approach to the effective management of
credit risk. To accomplish these employs machine learning techniques, specifically the KNN algorithm,
to analyze credit histories, seeking instances of both favorable and unfavorable credit. Extensive data
preparation and analysis are required as part of the methodology. The optimal K-value is determined
through accuracy testing after one- or two-way tests (such as ANOVA and Chi-Square) are conducted
to determine the statistical significance of independent variables. To overcome the obstacles posed by
high volatility, complex seasonality, and variation, the proposed method implements a robust cash flow
forecasting framework [13]. An ensemble technique is implemented following linear seasonal correction
and outlier smoothing, which integrates deep time series models and statistical models. A cross-learning
strategy is employed to train a single model using data from multiple series, thereby addressing the
issue of limited data. The approach surpasses the most advanced models that are presently accessible,
as evidenced by its validation on actual datasets.
3. Proposed Work
3.1. Data Collection and Processing
The initial step in the process of enhancing cash flow forecasting with AI and ML is the collection
and preparation of data. The initial phase involves the collection of all the requisite financial records
and reports from a variety of sources. This encompasses records and documentation regarding sales,
expenditures, cash flow, and external economic metrics. The prediction model can be enhanced and
understands the factors that influence cash flow can be enhanced by integrating data from multiple
sources. The data is prepared using a comprehensive method following data collection. This section
addresses potential disruptions to the forecasting model, including data gaps, inconsistencies, and
extreme cases. Data integrity is ensured by employing techniques such as normalization to scale various
categories of data and imputation to fill in absent values. Identifying and appropriately managing
anomalies, whether by eradicating or modifying them, is essential to prevent them from distorting the
results. Transforming and extracting features from data is also a component of the data preparation
process. Data organization is a critical component of preparing unprocessed data for use by machine
learning algorithms. Encoding techniques, such as one-hot encoding or label encoding, are employed
to convert category information into numerical representations.
Figure 1: System Architecture.
The predictive efficacy of the model is improved by the identification of essential data attributes
that have a substantial impact on cash flow through feature extraction. Principal Component Analysis
(PCA) and other dimensionality reduction techniques may assist in the elimination of extraneous
characteristics, thereby enabling the model to focus on important factors. The subsequent phase of
time series forecasting involves the alignment and sequencing of temporal data. It is essential to ensure
that the model accurately depicts the sequence of events by aligning data points with the appropriate
timestamps. This temporal congruence is essential for models such as Gated Recurrent Units (GRU)
and Long Short-Term Memory (LSTM) networks, which depend on sequential data patterns to generate
precise predictions. Figure 1 depicts Data Flow of the proposed system architecture.Data augmentation
is an additional essential element of preprocessing. In order to augment the quantity of data points,
this approach implements synthetic data generation and resampling. Additional data is incorporated
into the dataset to enhance the model’s generalizability and increase the accuracy of its predictions.
The data is partitioned into training, validation, and test sets during the penultimate stage of data
preparation. The training set is employed to train a machine learning model, the validation set to adjust
hyperparameters and prevent overfitting, and the test set to assess the model’s performance on new
data. Oracle Analytics and AI GPU environment have been used for Data preparation for creating ML
ops. One GB of ERP data has been sourced from Oracle ERP sandbox.
3.2. Feature Engineering
The application of feature engineering has the potential to significantly improve the accuracy of cash
flow forecasting models. These are currently in the process of identifying and developing supplementary
features from unprocessed data that have the potential to substantially impact cash flow predictions. A
model’s predictive potential may be enhanced by the transformation of unprocessed data into inputs
through the correct execution of feature engineering. The initial phase is to select pertinent attributes
by utilizing domain knowledge. These factors include sales data, payment due dates, seasonal patterns,
customer payment patterns, and macroeconomic variables such as interest and inflation rates. Each
specified characteristic assists the model in identifying trends and patterns by illuminating the factors
that affect cash flow. The model’s prediction accuracy is improved by the generation of additional
features that are generated in response to the identification of critical characteristics. This method
entails the development of time-based characteristics, including rolling statistics, delayed variables, and
moving averages, to accurately represent the temporal linkages of the data. In order to demonstrate
the influence of past events on future cash flows, delayed variables may be beneficial, and moving
averages may obscure short-term fluctuations while emphasizing trends over extended periods of time.
Furthermore, the relationships between various variables can be captured by implementing interaction
features. One method of understanding the impact of marketing on revenue and cash flow is to analyze
the correlation between marketing expenditures and sales volume. The model’s accuracy and complexity
may be further improved by the generation of polynomial features to represent non-linear interactions.
Each feature contributes equally to the model’s learning process, it is necessary to scale and normalize
them. The range of feature values is normalized using techniques such as Min-Max scaling or Z-score
normalization. This ensures that the model is not overly affected by the magnitude of any single feature.
In order to simplify the model, this employs dimensionality reduction techniques such as Principal
Component Analysis (PCA) to eliminate characteristics that are not essential. Principal component
analysis (PCA) enables us to maintain the model’s efficiency and clarity by retaining the most significant
features and eliminating the less significant ones.
3.3. Model Selection and Training
The selection and training of models are the crux of any dependable AI and ML system for financial flow
forecasting. The initial phase of this process involves the assessment of numerous machine learning
methodologies to determine which model most effectively captures the intricacies of cash flow dynamics.
Time series models, such as ARIMA and SARIMA, are appealing alternatives that can analyze sequential
data and identify patterns and seasonality. The prevalence of more sophisticated models is increasing
as a result of their superior capacity to capture complex patterns and long-term correlations in time
series data. Transformer networks, LSTM networks, and Gated Recurrent Units (GRU) are among
these models. After selecting the appropriate models, the subsequent step is to arrange the data for
training. The dataset comprises three distinct components: validation, training, and testing. The model
is instructed on the fundamental correlations and patterns present in the data by the training set. The
validation set must be employed to fine-tune the model’s hyperparameters in order to enhance its ability
to adapt to new data. Before the model is deployed into production, it undergoes a final evaluation on a
test set to ensure that its performance is unbiased. After numerous iterations of supplying the model
training data and fine-tuning its parameters, the process of predicting becomes effortless. In order to
achieve this, models such as LSTM and GRU employ backpropagation over time, a method that modifies
the weights of the neural network in accordance with the gradients of data errors. Dropout and early
ceasing are regularization techniques that are employed to prevent overfitting and maintain the model’s
accuracy when new, unknown data is introduced. It is crucial to modify the hyperparameters of the
model during the training procedure. In contrast to model parameters, hyperparameters are established
prior to the commencement of the learning process and regulate the entire training process. Some
of the strategies employed to identify the hyperparameters that optimize the model’s performance
include grid search, random search, and Bayesian optimization. Following the extensive testing of
various hyperparameter combinations and the measurement of the model’s output on the validation
set, the optimal configuration is determined. The model is subjected to a stringent test on the test
set following training and tuning in order to evaluate its accuracy and reliability. The model’s ability
to predict outcomes is evaluated using performance metrics such as MAE, RMSE, and MAPE. Good
models are capable of accurately predicting future cash flows by comprehending the subtleties of these
patterns, which is advantageous for long-term financial planning. The initial step in the development
of a dependable cash flow forecasting system is the selection and training of the appropriate models.
3.4. Algorithm
The technique utilizes a state-of-the-art machine learning model, such as an LSTM network, to forecast
future currency flows. LSTM is an ideal candidate for time series forecasting, as it has the capacity to
identify intricate temporal patterns and long-term relationships. This algorithm is capable of processing
data in sequences due to its ability to recall the impact of its prior states on its predictions. LSTM
networks are composed of linking layers of "neurons," or nodes, where each "node" represents a
mathematical operation. The input gate, the neglect gate, and the output gate comprise the majority
of this network’s architecture. The input gate regulates the quantity of new data that is incorporated
into the cell state, the neglect gate determines which data should be eliminated, and the output gate
determines the purpose for which the cell state should be used when it is updated. In order to acquire
knowledge from data sequences over extended periods of time, long short-term memory (LSTM)
neurons may selectively retain and neglect specific pieces of information. This is facilitated by their
gating mechanism. The LSTM algorithm trains itself through backpropagation through time (BPTT)
by iteratively modifying its internal weights. Computing the gradient of the loss function is a critical
phase in this process, as it quantifies the discrepancy between the predicted and actual income flows. In
order to mitigate this loss to the greatest extent possible, the algorithm modifies its weights to generate
more accurate predictions. Regularization techniques, including dropout, are implemented to mitigate
overfitting and guarantee that the model functions optimally with new data.
3.5. Hyperparameter Optimization
Hyperparameter tuning is a critical component of cash flow forecasting that employs sophisticated
algorithms, including Long Short-Term Memory (LSTM) networks, to enhance the efficacy of machine
learning models. Hyperparameters are immutable variables that regulate the learning process, while
parameters of a model are acquired during the training process. For optimal model performance, it
is imperative to configure them appropriately. The objective of hyperparameter optimization is to
achieve the highest level of accuracy and efficiency by conducting a systematic search for the optimal
combination of hyperparameters. Key hyperparameters for long short-term memory (LSTM) networks
include the learning rate, batch size, dropout rate, unit count per layer, and layer count. The learning
and generalizability of the model can be altered by adjusting these hyperparameters. Grid search is
frequently employed in hyperparameter optimization. It is necessary to specify a range of values for each
hyperparameter and subsequently execute the procedure with every conceivable combination of these
values. The most accurate combination of hyperparameters can be identified by comparing the model to
a validation set. Grid search encompasses the entire hyperparameter space, despite its computationally
intensive nature. Random search is an alternative to grid search that employs a predetermined range to
sample combinations of hyperparameters. While random search is less comprehensive than grid search,
it may often identify optimal or near-optimal settings more quickly, particularly in high-dimensional
areas. Bayesian optimization, a more sophisticated approach, employs a probabilistic model to simulate
the operation of various hyperparameter permutations. Using previous results as a foundation, this
method iteratively seeks hyperparameter spaces that are most likely to deliver optimal performance. In
order to identify the optimal hyperparameters, a significant amount of experimentation is required;
however, Bayesian optimization can be of assistance in this quest. The model’s cash flow projections
are rendered more precise and efficient by optimizing its learning process through hyperparameter
adjustments. To employ cash flow forecasting, an organization’s infrastructure must include a machine
learning model. During this phase, the forecasting model is integrated with the organization’s current
financial system to enable it to acquire real-time data and offer valuable insights. The initial phase of
data integration involves the development of a data conduit that would establish a connection between
the forecasting model and a variety of data sources. Real-time and historical financial data, as well
as data from an ERP system, accounting software, or external economic database, can be accessed
by establishing the appropriate connections. The pipeline ensures that data is consistently accurate
and current by enabling its continuous, error-free integration into the model. The second step after
configuring the data pipeline is to deploy the ML model to a production environment. It is a standard
practice to deploy the model on a server or cloud platform in order to ensure that it can receive data
inputs continuously and make predictions. The effective deployment of a model in the management
of real-time processing is contingent upon the optimization of computer resources and performance.
As part of the integration process, this offers an API that other components of the company’s system
may utilize to communicate with the forecasting model. The model’s ability to communicate with
other applications through the API facilitates a diverse array of capabilities, such as interaction with
reporting tools, automated data retrieval, and prediction updates. This compatibility guarantees that
the predictions’ outcomes are accessible for financial planning and decision-making. The objective of
creating user interfaces and displays is to facilitate the comprehension and utilization of prediction
results. Two functions of the diagrams and charts that are frequently used in these interfaces are the
ability to identify patterns and outliers, as well as the prediction of cash flow. Stakeholders can rapidly
comprehend the outcomes and make informed decisions when they have easy access to prognostic
data. Lastly, the integrated system must be consistently monitored and maintained to guarantee its
continued operation and reliability. It is regularly updated and refined to remain in accordance with the
evolving demands of the firm, data sources, and forecasting standards.
4. Results
The dataset that contains comprehensive financial and economic data obtained from many sources is
used to evaluate the proposed AI- and ML-based cash flow forecasting method. This series comprises
sales data, financial flow statements, expenditure reports, and real-time economic indicators. A robust
foundation for training and evaluating the forecasting models is provided by the dataset, which spans
five years. 60,000 data points are included in the sales records, which are derived from monthly
transactions. Expense reports, which contain 20,000 data points, provide a comprehensive account of
expenditures during the same time frame. The cash flow statements, which contain 20,000 pieces of
data, are accessible to the public every three months. Inflation and interest rates are merely two of the
1,825 data points that are accumulated daily to monitor real-time economic indicators. 30,000 data
points are also used to document the method of payment for each client monthly. The final dataset was
composed of 130,000 treated data points after data preprocessing addressed missing values, standardized
the data, and manufactured features. Delayed variables, interaction characteristics, moving averages,
and encoded categorical variables are all critical components. Table 1 depicts the dataset information.
Table 1
Dataset Information
Data Source Time Period Total Data Points
Sales Records 5 years 60000
Expense Reports 3 years 20000
Cash Flow Statements 5 years 20000
Economic Indicators Real time 1825
Customer Payment Behavior 2 years 30000
A variety of critical parameters are employed to evaluate the proposed forecasting method. Some of
these metrics include the MAE, RMSE, and MAPE. The metrics presented below may facilitate a more
comprehensive comprehension of the LSTM model’s cash flow forecasts’ reliability. The MAE is a metric
that is used to evaluate the typical magnitude of forecasting errors, excluding the directional component
of these errors. The model’s efficacy is enhanced as the MAE decreases. The average squared difference
between the predicted and actual values can be determined using the root-mean-squared error (RMSE)
measure. Its penalties for more substantial errors are more severe than those of MAE. Mean Absolute
Percentage Error (MAPE) is a normalized metric that quantifies forecast accuracy as a percentage. The
metrics as output of the LSTM model in this investigation were from Table 2:
Table 2
LSTM model
Metric Value
Mean Absolute Error (MAE) $900
Root Mean Squared Error (RMSE) $1200
Mean Absolute Percentage Error (MAPE) 4.2%
Contrasting the proposed LSTM-based method with more conventional forecasting techniques, such
as ARIMA, is the sole method by which the outcomes can be evaluated. The performance metrics of the
two methods are summarized in the following Table 3 and Figure 2:
Table 3
Comparison of proposed and existing methods
Method MAE ($) RMSE ($) MAPE
ARIMA 1200 1500 5.5%
LSTM 900 1200 4.2%
GRU 850 1150 4.0%
Transformer 800 1100 3.8%
Figure 2: Model Comparison graph.
The LSTM model outperforms ARIMA in terms of prediction accuracy and reliability, as evidenced by
these three metrics. The proposed method significantly reduces MAE and RMSE, while simultaneously
halving MAPE in comparison to the current state. In cash flow forecasting, advanced models such
as GRU and Transformer networks outperform LSTM, illustrating the ongoing evolution of machine
learning approaches. The findings support the notion that AI-powered technology is capable of more
accurately predicting future cash flows than conventional methods, which could be beneficial for
financial decision-making and strategic planning. Figure 3 depicts the time series plot of historical and
forecasting cash flow data. Figure 4 depicts the training and validation loss over epoch.
5. Conclusion
AI and ML-based cash flow forecasts have the potential to revolutionize financial management by
producing more precise and dependable projections. Businesses can generate dynamic and precise
Figure 3: Time Series Plot.
Figure 4: Training and Validation Loss.
forecasts by employing sophisticated models such as LSTM networks and modifying hyperparameters
through methods such as grid search or Bayesian optimization. Organizations can enhance their
financial stability and make more informed decisions by leveraging the models’ robust visualization
capabilities, real-time data processing, and ease of interaction with existing financial systems. This
innovative approach not only improves prediction accuracy but also aids resource allocation, strategic
planning, and overall operational efficiency. As a result, it fosters competitiveness and expansion in a
dynamic industry.
Declaration on Generative AI
The author(s) have not employed any Generative AI tools.
References
[1] A. Alsuwailem, A. Saudagar, Anti-money laundering systems: a systematic literature re-
view, Journal of Money Laundering Control 23 (2020) 833–848. URL: https://doi.org/10.1108/
JMLC-02-2020-0018. doi:10.1108/JMLC-02-2020-0018.
[2] K. R. Dalal, M. Rele, Cyber security: Threat detection model based on machine learning algorithm,
in: 2018 3rd International Conference on Communication and Electronics Systems (ICCES), 2018,
pp. 239–243.
[3] K. D. Rohit, D. B. Patel, Review on detection of suspicious transaction in anti-money laundering
using data mining framework, 2015.
[4] E. L. Paula, M. Ladeira, R. N. Carvalho, T. Marzagão, Deep learning anomaly detection as support
fraud investigation in brazilian exports and anti-money laundering, in: 2016 15th IEEE International
Conference on Machine Learning and Applications (ICMLA), 2016, pp. 954–960.
[5] A. Faccia, N. R. Moçteanu, L. P. L. Cavaliere, L. J. Mataruna-Dos-Santos, Electronic money launder-
ing, the dark side of fintech: An overview of the most recent cases, in: ICIME 2020: Proceedings
of the 2020 12th International Conference on Information Management and Engineering, ACM
International Conference Proceeding Series, Association for Computing Machinery, 2020, pp.
29–34.
[6] M. Rele, D. Patil, Intrusive Detection Techniques Utilizing Machine Learning, Deep Learning, and
Anomaly-based Approaches, 2023 IEEE International Conference on Cryptography, Informatics,
and Cybersecurity (ICoCICs) (2023) 88–93.
[7] M. Tiwari, A. Gepp, K. Kumar, A Review of Money Laundering Literature: The State of Research
in Key Areas, Pacific Accounting Review 32 (2020) 271–303.
[8] R. Al-Shabandar, G. Lightbody, F. Browne, J. Liu, H. Wang, H. Zheng, The Application of Artificial
Intelligence in Financial Compliance Management, in: AIAM 2019: Proceedings of the 2019
International Conference on Artificial Intelligence and Advanced, volume 8 of ACM International
Conference Proceeding Series, Association for Computing Machinery, 2019, pp. 1–6.
[9] M. E. Lokanan, Data mining for statistical analysis of money laundering transactions, Journal of
Money Laundering Control 22 (2019) 753–763.
[10] M. Rele, D. Patil, Supervised and Unsupervised ML Methodologies for Intrusive Detection in Nu-
clear Systems, 2023 International Conference on Network, Multimedia and Information Technology
(NMITCON) (2023) 1–7.
[11] A. Khudhur, A. Al-Alawi, The Use of Machine Learning to Forecast Financial Performance: A
Literature Review, 2024, pp. 1–6. doi:10.1109/ICETSIS61505.2024.10459393.
[12] M. Becha, O. Dridi, O. Riabi, Y. Benmessaoud, Use of Machine Learning Techniques in Financial
Forecasting, 2020, pp. 1–6.
[13] Y. Zheng, K. Tu, A Robust Forecasting Framework for Multi-Series Cash Flow Prediction, 2023, pp.
898–902. doi:10.1109/ICICSP59554.2023.10390581.