Open Access

Manuel MuthORCiD, Jonas Engel, Gerd NuferORCiD

• While big data settings are prominent in AI forecasting literature, many real-world market applications face the opposite challenge: generating accurate forecasts from small datasets with fewer than 500 observations. This paper compares pretrained AI models with established forecasting benchmarks under small-data conditions. In an experiment, new industrial orders are forecasted multiple steps ahead (h=1,…,12) based on systematically varied data lengths n {6, 12, 24, 48, 96, 192, 384}. Specifically, using actual industrial order data from German market, pretrained AI models (TabPFN, Chronos) are compared with established time-series models (ARIMA, ETS, Theta) as well as non-pretrained AI forecasting methods (Random Forest, XGBoost, LSTM). The empirical results suggest that, within this specific industrial data context, the pretrained AI model TabPFN performs best on average, followed by Chronos. Non-pretrained AI methods achieve comparable results, specifically LSTM and Random Forest, while established time-series models and the Naive baseline show higher performance errors in this setting. This paper provides context-specific evidence that pretrained AI models may expand the toolkit of AI-driven market analytics by reducing model tuning requirements and data prerequisites, with implications for resource-constrained organizations operating in data-scarce environments.