上述結果顯示，小型開源大型語言模型是可行的零樣本與少樣本外匯預測工具，且在未經微調的情況下，提示策略對預測表現的影響較新聞資料增強更為可靠。;Three LLMs (Llama3-8B， Mistral-7B， Qwen2.5-7B) are evaluated with the ChatTime tokenization framework on five USD currency pairs (IDR， TWD， AUD， EUR， SGD) across three horizons (1， 5， and 10 days)， under a 2 x 2 factorial ablation design， against seven baselines (ARIMA， LSTM， XGBoost， iTransformer， vanilla Transformer， Naive， Moving Average). Performance is measured with RMSE， MAE， MAPE， Directional Accuracy (DA)， and R^2.
LLMs outperform all baselines across every currency pair and horizon， with Qwen2.5-7B achieving the lowest mean RMSE of 27.48， compared to 91.81 for the best baseline (XGBoost). A Diebold--Mariano test confirms lower LLM loss in all 45 model-currency-horizon comparisons， significant in 38. All LLMs achieve positive mean R^2 (up to 0.50)， while every baseline achieves negative mean R^2. News augmentation does not consistently improve performance， with higher RMSE observed in 32 of 45 units when news is added and the effect significant in only 11; no significant DA effect is found. Few-shot prompting reduces mean RMSE by approximately 8.2% and raises mean DA from 51.79% to 53.25% relative to zero-shot， reaching significance in 16 of 45 units. A pattern consistent with a disordinal interaction between news and prompting strategy is observed for DA， suggesting the effect of news augmentation depends on the prompting approach used.
These findings demonstrate that small open-source LLMs are viable zero-shot and few-shot forecasters for FX markets， and that prompting strategy is a more reliable performance driver than news augmentation in the absence of fine-tuning.
