Machine Learning-Based Forecasting of Business Performance Under Market Uncertainty

Abstract

Forecasting firm-level performance becomes markedly harder when the economic environment shifts from a stable to a volatile regime, yet most machine-learning (ML) benchmarks evaluate models under a single, homogeneous data condition. This paper proposes an evaluation protocol for performance forecasting under market uncertainty and applies it in a controlled simulation study. A synthetic panel of 60 firms observed over 80 quarters is generated from an autoregressive data-generating process in which a common macro factor and idiosyncratic drivers determine performance, and in which the innovation variance triples at the midpoint of the sample to mimic an uncertainty shock. Five forecasters, namely a naive last-value rule, ridge regression, random forest, gradient boosting and a two-layer multilayer perceptron, are trained and evaluated separately in the low- and high-uncertainty regimes, and a quantile gradient-boosting model is used to assess whether prediction intervals retain their nominal coverage after the regime change. Root-mean-square error rises by a factor of roughly 2.6 to 3.3 across all learners when uncertainty increases, the ranking of learners changes between regimes, and the nominal 80% interval covers only 71.5% of realised outcomes in the volatile regime. The results show that the gains from flexible learners are regime-dependent, that interval calibration degrades precisely when decision-makers need it most, and that forecasting systems should report regime-conditional accuracy and recalibrate uncertainty estimates as part of their operating procedure. The study contributes a reproducible protocol and a set of design principles for deploying ML forecasters in uncertain markets.