Collaboration with Bankomat
| Title | Collaboration with Bankomat |
|---|---|
| Summary | Explainable AI for Forecasting in Corporate Environments ( RQ: How to provide interpretable/explainable forecasts that align with business logic? |
| Keywords | |
| TimeFrame | Fall2025 |
| References | |
| Prerequisites | |
| Author | |
| Supervisor | Zahra Taghiyarrenani, Parisa?, Slawomir?, Sepideh? |
| Level | Master |
| Status | Open |
Explainable AI for Forecasting in Corporate Environments
We propose two master’s thesis projects in collaboration with Bankomat AB, focusing on the development of interpretable and explainable forecasting systems for ATM cash-demand prediction. Following, we describe how each of these thesis projects can be defined, representing two complementary directions.
Background and Motivation Time-series forecasting plays a crucial role in financial and operational decision-making, where organizations must anticipate future demand, costs, or risks based on historical patterns. For an ATM network operator such as Bankomat AB, accurate cash-demand forecasting ensures that every machine has enough cash to meet customer needs while minimizing costly over-supply and transport inefficiencies. However, as forecasting models become increasingly data-driven and complex, they tend to operate as “black boxes.” This lack of transparency makes it difficult for decision-makers to understand why a model predicts higher withdrawals at a certain time or which factors drive a sudden change. Therefore, developing forecasting systems that are not only accurate but also explainable is essential for building trust, ensuring accountability, and aligning predictions with real-world business logic. Two promising directions for achieving this are:(1) Enhancing Transparency through Explainable AI in Forecasting, (2) leveraging Large Language Models for Contextual and Human-Centric Explanations.
Thesis1: Enhancing Transparency through Explainable AI in Forecasting Explainable AI methods aim to uncover how input variables and temporal patterns influence forecasting outcomes [1, 2]. In the context of ATM cash-demand prediction, such methods can reveal which recent time window, seasonal pattern, or external variable (e.g., weekday, salary period, or temperature) most affects the forecast. Techniques such as SHAP, attention visualization, and prototype learning can quantify these effects and visualize the model’s reasoning, helping analysts trace how each decision is formed. Research Question:
How can we design an efficient data-driven ATM cash-forecasting model that remains transparent by identifying and visualizing the most influential temporal and contextual factors behind each prediction?
This study will further explore how heterogeneous data sources, such as transaction history, holidays, temperature, and salary periods, can be integrated into an interpretable forecasting framework that combines predictive accuracy with human-understandable reasoning.
Thesis 2: Leveraging Large Language Models for Contextual and Human-Centric Explanations Large Language Models (LLMs) offer a complementary approach to explainability by reasoning jointly over numerical and textual information [3, 4]. In ATM cash forecasting, LLMs can interpret external signals, such as news headlines, event reports, or policy announcements, to provide narrative explanations that connect real-world context with model predictions. Inspired by recent reflective forecasting frameworks [5], LLMs can dynamically integrate event information and refine their reasoning when prediction errors reveal missing contextual factors. This enables the generation of natural-language justifications, for example, explaining a forecasted increase in withdrawals as a result of a salary day coinciding with a regional festival. Research Question:
How can LLMs be used to generate accurate and human-interpretable explanations for ATM cash-demand forecasts by incorporating both numerical data and external contextual information?
This study will further investigate how reflection-based reasoning mechanisms can improve the clarity, consistency, and business alignment of such explanations, supporting more transparent and trustworthy decision-making for Bankomat cash management operations.
Both theses aim to advance the transparency and practical usefulness of data-driven forecasting systems in financial operations. The XAI-based project is expected to deliver a forecasting framework that not only achieves high predictive accuracy but also provides clear visual and quantitative explanations for the factors influencing ATM cash demand. The LLM-based project, in turn, will develop a system that incorporates textual information, such as news, holidays, and regional events, to not only enhance forecasting accuracy but also generate meaningful insights into the underlying causes of demand fluctuations.
References: 1. Arsenault PD, Wang S, Patenaude JM. A survey of explainable artificial intelligence (XAI) in financial time series forecasting. ACM Computing Surveys. 2025 May 7;57(10):1-37. 2. Jahin MA, Shahriar A, Amin MA. MCDFN: supply chain demand forecasting via an explainable multi-channel data fusion network model. Evolutionary Intelligence. 2025 Jun;18(3):66. 3. Tang H, Zhang C, Jin M, Yu Q, Wang Z, Jin X, Zhang Y, Du M. Time series forecasting with llms: Understanding and enhancing model capabilities. ACM SIGKDD Explorations Newsletter. 2025 Jan 22;26(2):109-18. 4. Schoenegger P, Park PS, Karger E, Trott S, Tetlock PE. Ai-augmented predictions: Llm assistants improve human forecasting accuracy. ACM Transactions on Interactive Intelligent Systems. 2025 Feb 10;15(1):1-25. 5. Wang X, Feng M, Qiu J, Gu J, Zhao J. From news to forecast: Integrating event analysis in llm-based time series forecasting with reflection. Advances in Neural Information Processing Systems. 2024 Dec 16;37:58118-53.
