Which model should you trust?
Research question: which modelling approach (VAR/VECM vs. LSTM) most reliably forecasts near-term (1–20 day) changes in the Canadian 10-year minus 2-year yield spread, balancing forecast accuracy, stability, and interpretability?
Current verdict: VECM
At the 5-day horizon — the midpoint of the operational 1–20 day window — VECM has the lowest RMSE (0.0708) among the multivariate/ML candidates, an improvement of 0.0% over the random-walk benchmark.
Trade-offs: VECM is directly interpretable (coefficients map onto specific lagged relationships between rates, yields, USD/CAD, and CPI), which matters for a decision-support context where analysts need to explain *why* a forecast moved, not just that it moved. Its advantage is not consistent across all three horizons — treat forecasts beyond 5 days with more caution.