Performance of Single-Layer Transformers in Stock-Price Trend Forecasting

Accurate forecasting of financial time series demands models that capture long-range dependencies while remaining transparent and computationally tractable. In this study, we introduce a single-layer Transformer encoder for forecasting daily closing prices of Nasdaq equities and propose a segment-based performance evaluation framework. Each equity's historical price series is partitioned into contiguous training and testing intervals, discarding segments that fall below a minimum length while normalizing all intervals independently. Within each interval, fixed-length sliding windows are augmented with sinusoidal positional embeddings and processed by a Transformer block comprising a linear projection, multi-head self-attention with residual connections, layer normalization and dropout, followed by a two-stage feed-forward network and global average pooling to yield point forecasts. Models are trained with early stopping on a held-out validation split. Forecast performance is quantified per interval using the Pearson correlation coefficient and Mean Squared Error (MSE). Every interval-level correlation coefficient-one for each equity-interval pair-is aggregated into histograms, with interquartile-range filtering applied to emphasize central tendencies. Empirical results across dozens of tickers demonstrate that this minimal architecture reliably captures directional trends under volatile market conditions. The single-layer design enhances interpretability and computational efficiency, establishing a reproducible baseline for future deep-learning research in financial time-series forecasting.