# Internship - Predicting confidence bands for future account balances at ING ING is hiring an Internship - Predicting confidence bands for future account balances. Level rates it Builds AI ●●●●; you can [apply on ING](https://jobsbylevel.com/go/4ac3918f-0cce-47c9-a7af-9626f082099a?utm_medium=ai-agent). AI Level 4, AI centrality 85 out of 100. CDR (Amsterdam - Cedar). ## Details - Company: [ING](https://jobsbylevel.com/companies/ing) - AI level: AI Level 4 (score 85 out of 100) - Location: CDR (Amsterdam - Cedar) - Posted: October 29, 2026 - Apply: https://jobsbylevel.com/go/4ac3918f-0cce-47c9-a7af-9626f082099a?utm_medium=ai-agent - Free CV review for this job: https://jobsbylevel.com/cv-review?job=internship-predicting-confidence-bands-for-future-account-balances-at-ing-e562e1 ## Prepare for this job The job page lists 5 likely interview questions for this offer, with a CV checklist: https://jobsbylevel.com/jobs/internship-predicting-confidence-bands-for-future-account-balances-at-ing-e562e1 ## Description There are many standard models and methodologies available that can provide the usual point predictions when it comes to time series forecasting, ranging from classical (linear) models to machine learning models to foundation models. However, when not only point predictions are desired, but also a confidence or uncertainty range, then finding a suitable solution becomes much more difficult. Many standard models, such as the bulk of machine learning models, do not provide a way to learn and predict confidence intervals out of the box. Classical models often do provide confidence levels but are generally limited in the types of times series that they can describe well. And the recent foundation models for time series forecasting, created on the success of Large Language Models, may come with the ability to forecast confidence levels, but are still immature and largely unproven. At ING Wholesale Banking Advanced Analytics we research possibilities to send account managers alerts when their clients have exceptionally high or low account balances. For this purpose, we do not so much need to predict the balance itself (point predictions) but rather predict an interval in which the balance is expected to lie. So that an alert is warranted when the balance exceeds this interval. In previous research we focused primarily on (machine learning) models with an inherent capability to produce intervals. We like to continue the research, where we are also interested in methods that may produce intervals for already trained point prediction models, such as conformal predictions. The primary point for this particular thesis project is the reliability of the intervals, at least in the marginal sense. The next is getting small enough intervals to be useful for alerting. Context Account managers in ING Wholesale Banking handle dozens of large corporate clients. For each client they need to be aware of their client’s financial status and movements, so they can for example check if mutual agreements are still upheld, or if there is opportunity for up- or cross-selling. If a client takes on an abnormally large or small account balance position, then this can be an alert for the account manager to take a closer look at the client. To automate these alerts, the WBAA department is modeling abnormal behavior, by first modeling the normal expected range of the client’s balance development and then determining if the actual balance significantly deviates out of this range. Hence the desire for confidence bands on time series predictions of account balances. Research goal We are agnostic to the model or methodology that will be researched, but given the diversity of the account balances and given the past research that we have done, we do believe that classical models, such as ARIMA and linear state space models, are insufficient for this task. We have a number of potential solutions in mind: machine learning + conformal predictions , machine learning + block bootstrap sampling of residuals, training a neural network with a double pinball loss , or simply trying foundation models , but it is up to the student (+ supervisor) to choose and focus on one or more methods. The foreseen end-result is a set of measures that state how well the chosen model(s) can forecast reliable confidence or uncertainty intervals and how accurate they are in comparison with a set of baseline models (comparing widths of intervals). This will include: Assessing potential solutions via literature research and/or short experiments and ultimately deciding on one or more solutions to understand, implement and test Implementing the chosen solutions as well as one or more baseline methods (e.g. auto-ARIMA, simple bootstrapping, out-of-the-box deep learner) Choosing metrics that measure 1) the reliability of the produced intervals and 2) the width of the produced intervals Implementing a pipeline to be able to test and compare solutions out-of-time Analyzing for which balance behaviors The description is cut here. Read the full offer: https://jobsbylevel.com/jobs/internship-predicting-confidence-bands-for-future-account-balances-at-ing-e562e1 Source: https://jobsbylevel.com/jobs/internship-predicting-confidence-bands-for-future-account-balances-at-ing-e562e1 ## Cite this page Level. https://jobsbylevel.com/jobs/internship-predicting-confidence-bands-for-future-account-balances-at-ing-e562e1. Get job alerts: https://jobsbylevel.com/newsletter