When Forecasting Errors Have a Price: Macroeconomic Prediction under Asymmetric Loss

Témavezető:	Sebestyén Géza
email:	sebestyen.geza@mcc.hu

Projekt leírás

Description of the Research Topic:

Forecasts are usually trained and evaluated with symmetric measures such as mean squared error, which implicitly assume that overprediction and underprediction of the same magnitude are equally costly.

In many real decisions this assumption is unrealistic. Underestimating inflation may leave a portfolio insufficiently hedged, overestimating tax revenue may create an unexpected budget deficit, while underestimating demand may lead to costly capacity shortages.

The project investigates how forecasts change when the model is optimized for an explicitly asymmetric economic loss function rather than for conventional statistical accuracy alone. The methodological spectrum may include quantile loss, asymmetric quadratic loss, LINEX loss, and a custom piecewise differentiable cost function calibrated to a concrete decision problem.

Using the same inflation, GDP, employment, tax-revenue, sales, or energy-demand data, the student will compare a symmetric benchmark with cost-sensitive regression, XGBoost, or a simple neural network. Rolling-window backtesting and sensitivity analysis across several cost ratios will show when a deliberately shifted forecast reduces expected economic loss and how much conventional accuracy must be sacrificed to achieve that reduction.

The final output may include reproducible Python code and a configurable decision-support dashboard in which users select the relative cost of positive and negative errors, inspect the resulting optimal forecast, and compare statistical performance with realized or simulated monetary loss.

By choosing this topic, the student will acquire highly transferable and marketable skills in time-series forecasting, optimization, probabilistic modelling, machine learning, custom loss-function design, model validation, and decision-support development, opening up career opportunities in central banking, public policy, banking, insurance, asset management, consulting, energy, retail, logistics, manufacturing, pharmaceuticals, and any organization in which forecasting errors have measurable and direction-dependent consequences.

Skills and Knowledge to Be Acquired:

The project will provide the student with an opportunity to progressively learn and apply:

• the mathematical foundations of time-series forecasting and statistical decision theory;
• the distinction between statistical forecast accuracy and the economic value of a forecast;
• symmetric and asymmetric loss functions, including squared error, quantile, LINEX, and custom piecewise losses;
• the derivation or numerical computation of an optimal point forecast under a specified loss function;
• Python-based data analysis and modelling, particularly using pandas, statsmodels, scikit-learn, XGBoost, and optionally PyTorch;
• the collection, cleaning, transformation, and temporal alignment of macroeconomic or business data;
• rolling-window backtesting and methods for preventing temporal data leakage;
• the implementation of custom objectives and gradients in machine-learning models;
• the evaluation of forecasts using both conventional error metrics and decision-specific monetary loss;
• sensitivity analysis for alternative cost ratios, forecast horizons, and economic regimes;
• the development of a configurable forecasting dashboard, for example using Streamlit;
• the interpretation of English-language methodological literature and the presentation of technical findings to non-technical decision-makers.

The project can be completed at several levels of difficulty. Its core version would compare a symmetric benchmark with quantile, or LINEX-based forecasts for one selected variable. An extended version could estimate cost parameters from observed decisions, implement a differentiable custom loss in XGBoost or a neural network, connect predictions to a downstream optimization problem, and test whether decision-focused training produces statistically significant reductions in economic loss.

Hivatkozások

• Christoffersen, P. F. and Diebold, F. X. (1997): “Optimal Prediction under Asymmetric Loss.” Econometric Theory, 13(6), 808–817. link
• Elliott, G., Komunjer, I. and Timmermann, A. (2005): “Estimation and Testing of Forecast Rationality under Flexible Loss.” The Review of Economic Studies, 72(4), 1107–1125. link
• Donti, P., Amos, B. and Kolter, J. Z. (2017): “Task-Based End-to-End Model Learning in Stochastic Optimization.” Advances in Neural Information Processing Systems, 30. link
• Wilder, B., Dilkina, B. and Tambe, M. (2019): “Melding the Data-Decisions Pipeline: Decision-Focused Learning for Combinatorial Optimization.” Proceedings of the AAAI Conference on Artificial Intelligence, 33(1), 1658–1665. link