← back to mock pineapple
Mock Pineapple · Forecasts vs Reality
One backtest, three models, one ensemble
Cutoff: 2024-05-31. The system was trained on data up to that Friday, then asked to forecast 30 business days forward. Each pair below shows the actual realized rate (gray, solid) and what each of the three models — plus the horizon-adaptive ensemble — predicted. Pick a pair to switch.
SARIMA uses each pair's tuned (p,d,q)(P,D,Q,s) from tuned_hyperparameters.json. Prophet runs with weekly + yearly seasonality. LightGBM has 38 features (price lags + rolling stats + calendar + 5 macro signals: DXY, VIX, US 10Y, gold, oil) and trains a separate model per horizon. The ensemble weights are inverse-squared MAPE, normalized — so the model that's been most accurate gets the biggest vote at each pair.
What you're looking at: a single 30-day backtest. The gray line is the actual rate. Each colored line is one model's 30-day forecast made at the cutoff (2024-05-31). The purple ensemble line is the inverse-MAPE-weighted blend — typically the closest to gray. Some models miss obvious peaks because point forecasts are smooth by construction; the ensemble's job is to average out individual biases.