Monetary-policy announcements and central-bank communications play a central role in foreign exchange markets, yet their qualitative, unstructured form makes their forecasting value difficult to quantify. While prior research has largely focused on sentiment extracted from financial news, comparatively little is known about the relative contribution of different dimensions of monetary-policy communication. Existing studies primarily evaluate whether textual information improves overall forecasting performance but provide limited insight into which communication channels drive such improvements. To address this gap, this paper introduces a statistical attribution methodology that decomposes monetary-policy communication into interpretable channels and quantifies their incremental forecasting contribution under false-discovery-rate control. Monetary-policy news is transformed into structured communication signals using large language models (LLMs) and temporal feature engineering. These signals are evaluated using rolling-window experiments with tree-based machine-learning models. The results show that monetary-policy communication contains measurable predictive information. Attribution analysis shows that predictive value is concentrated in a small subset of signals, with communication timing providing the strongest individual feature-level contribution, targeted communication-activity measures also contributing positively, and LLM-derived sentiment providing complementary information at the group level. The findings indicate that communication-based forecasting value extends beyond sentiment alone and that attribution, rather than aggregate accuracy alone, is central to evaluating news-derived signals.
Figures & tables
Fig. 1 : Overview of the proposed monetary-policy communication forecasting and attribution workflow.
TABLE I : Feature categories used in the forecasting methodology.
Horizon
Best ML Model
RW
R2
NRMSE
R2
NRMSE
1h
0.9150
0.00549
0.9988
0.00065
3h
0.9039
0.00584
0.9967
0.00109
8h
0.8886
0.00628
0.9916
0.00173
24h
0.8442
0.00742
0.9764
0.00289
TABLE II : Benchmark comparison for level forecasting.
Fig. 2 : Mean RMSE across the cumulative ablation sequence at the 1-, 3-, 8-, and 24-hour horizons. Lower values indicate better forecasting performance. Step 0 is the full feature set; each subsequent step removes one feature.
H
Pinball Loss
WIS
Cov90 (%)
Full
Step 7
Full
Step 7
Full
Step 7
1h
0.003329
0.001479
0.002394
0.001062
53.72
68.92
3h
0.003731
0.001989
0.002682
0.001429
63.23
58.47
8h
0.004001
0.002829
0.002877
0.002032
61.85
50.88
24h
0.004600
0.004102
0.003308
0.002946
57.85
46.29
TABLE III : Full-feature versus Step 7 probabilistic forecasting performance using uncalibrated quantile forecasts. Lower pinball loss and WIS indicate better probabilistic performance; Cov90 reports empirical coverage of the nominal 90% interval before conformal calibration.
Fig. 3 : Step 7 LightGBM quantile forecasts at the 1- and 24-hour horizons. Black shows the observed USD/CAD exchange rate, red the median forecast, and shaded regions the 50% and 90% prediction intervals. Responsiveness and empirical coverage decline at longer horizons.
Horizon
DA (%)
Majority (%)
Persistence (%)
Balanced Acc.
Macro F1
AUC
1 hour
66.08
63.05
62.61
0.721
0.660
0.744
3 hours
65.12
61.54
61.57
0.698
0.651
0.729
8 hours
66.55
57.80
61.51
0.705
0.658
0.707
24 hours
63.22
51.14
57.70
0.626
0.604
0.634
TABLE IV : Directional forecasting performance relative to benchmark strategies.
Fig. 4 : Mean directional accuracy across the cumulative ablation sequence at the 1-, 3-, 8-, and 24-hour horizons. Higher values indicate better directional classification. Step 0 denotes the full feature set; each subsequent step removes one additional feature.
Feature
Mean ΔR2
Mean Δ DA
95% CI
Global FDR q -value
Positive contributors
is_news_hour
+0.012
+0.04 pp
[+0.006,+0.018]
0.003
fed_non_neutral_last_24h
+0.005
-0.01 pp
[+0.003,+0.008]
0.003
fed_articles_24h
+0.005
-0.10 pp
[+0.002,+0.008]
0.008
boc_news_last_24h
+0.005
-0.09 pp
[+0.001,+0.009]
0.041
boc_articles_24h
+0.003
+0.08 pp
[+0.001,+0.004]
0.026
TABLE V : FDR-supported feature-level attribution results. Positive ΔR2 indicates that feature removal reduces forecasting performance, whereas negative ΔR2 indicates improved performance after feature removal.
Feature Group / Representative Feature
Mean ΔR2
Mean Δ DA
Feature groups
News timing only
+0.0005
-0.01 pp
News volume only
-0.0018
-0.01 pp
LLM-derived sentiment only
+0.0020
+0.02 pp
Combined
+0.0001
-0.00 pp
Representative LLM-derived sentiment features
TABLE VI : Group-level attribution results and representative LLM-derived sentiment features.
†Centre for Artificial Intelligence Research, University of Agder, Norway · ‡Head of AI, Bernoly AG, Zurich, Switzerland · ¶La Jolla Private Wealth Group, Wells Fargo Advisors, La Jolla, CA, USA