Forecasting Macroeconomic Indicators with Transformer-Based Time-Series Models

It is important to be able to forecast macroeconomic indicators accurately for economic policy and financial decisions. Most conventional statistical models perform poorly in modeling complex patterns in high-dimensional, non-linear and long-range dependent financial time series. In this work we focus on the use of transformer-based architecture for time-series forecasting of macroeconomic indicators like GDP growth, inflation rates, unemployment etc. Furthermore, fatal combinations among a set of attributes can jeopardize the continued sales within the remaining set of attributes, and the products in such a deadly combination set are removed from further consideration. The transformer model based on the self-attention mechanism can dynamically encode a large-scale number of temporal dependencies and heterogeneous feature relationships, without the restriction of fixed-window observation and implicit order constraints. We fine-tune several transformer-based models, including vanilla Transformer, Temporal Fusion Transformer (TFT), and Informer, on publicly available macroeconomic datasets at quarterly and monthly granular resolutions. We conduct rigorous comparisons with autoregressive integrated moving average (ARIMA) and Long Short-Term Memory (LSTM) benchmarks and find that transformer models outperform these baselines in terms of mean absolute errors across the board and lead to more stable forecasts, especially under multi-step and turbulent economic settings. Ablation studies verify the necessity of positional encoding and multimodal attention for modeling economic seasonality and the exogenous shocks. These findings suggest that the transformer-based approach represents a powerful framework for time-series forecasting in economics: it provides not only a state-of-the-art accuracy, but also interpretability via attention weights as well as the flexibility to integrate related information such as policy changes or global events.

Ҳали таржима қилинмаган