7-day R² of 0.70 using Trade Imbalance Ratio and gradient-boosted models
HYPERION Freight Forecasting achieves 7‑day R² of 0.70 and MAE of 611 USD/TEU on Far East→US West Coast rates by leveraging a proprietary Trade Imbalance Ratio feature and time‑series validated gradient‑boosted models.
The system produces 7/14/30‑day forecasts, with 14‑day R² of 0.50 for operational planning and 30‑day directional guidance, all backed by a foundry of benchmarked algorithms and rigorous backtesting.
Provide lane‑specific short‑ and mid‑term freight rate forecasts that support procurement timing, contract negotiation, and risk hedging under volatile market regimes.
Focus on robust out‑of‑sample performance and explainability via feature importance and domain‑grounded indicators rather than opaque black‑box signals.
Far East→US West Coast and reciprocal lane (eastbound/westbound) for imbalance measures, daily since 2018.
Dry bulk index proxy and Brent fuel prices to capture capacity and cost dynamics across macro conditions.
Feature engineering: Compute FEUW/UWFE ratio and multi‑lag stacks; exclude contemporaneous target to prevent leakage.
Model foundry: Ridge, Random Forest, LightGBM, XGBoost, and CatBoost benchmarked per horizon with backtesting and holdout.
Selection: XGBoost champions 7‑day; LightGBM strong on 14‑day; ensemble diagnostics retained.
Key signals guiding the forecasting system:
Higher values precede rate spikes:
$$ \text{TIR} = \frac{\text{FEUW}}{\text{UWFE}} $$
Market and fuel lags transmit macro and cost pressures; lane‑specific lags capture momentum and mean‑reversion.
Tree ensembles capture nonlinear interactions among fuel, macro proxies, and lane imbalances while retaining feature attributions.
XGBoost champions 7‑day; LightGBM strong on 14‑day; ensemble diagnostics retained.
$$ R^{2} = 1 - \frac{\sum (y-\hat{y})^{2}}{\sum (y-\bar{y})^{2}} $$
Primary variance‑explanation metric
$$ \text{MAE} = \frac{1}{N}\sum |y-\hat{y}| $$
Procurement‑aligned error in USD/TEU
With lag features:
$$ \text{TIR}_{t} = \frac{\text{FEUW}_{t}}{\text{UWFE}_{t}} $$
df["tir"] = df["feuw"]/df["uwfe"]
lags = [1,7,14,30]
for col in ["feuw","uwfe","bdi_proxy","brent","tir"]:
for L in lags:
df[f"{col}_lag_{L}"] = df[col].shift(L)
features = [c for c in df.columns if "lag_" in c]X = df[features].iloc[:-7].dropna()
y = df["feuw"].shift(-7).iloc[:-7].loc[X.index]
model = xgb.XGBRegressor(max_depth=6, n_estimators=400, learning_rate=0.05)
model.fit(X, y)
yhat = model.predict(X_test)
r2 = r2_score(y_test, yhat); mae = mean_absolute_error(y_test, yhat)$$ R^{2}=0.70 $$, $$ \text{MAE}=611 $$ USD/TEU; 14‑day: $$ R^{2}=0.50 $$; 30‑day: directional guidance under higher uncertainty.
2,698 daily records processed with real‑time pipeline hooks and automated QC for production readiness.
Leakage controls are enforced; horizon‑specific champions avoid one‑size‑fits‑all degradation.
Capture nonlinear interactions among fuel, macro proxies, and lane imbalances while retaining feature attributions.
Live ingestion, automated serving, monitoring for regime breaks, and alerting when predicted spikes exceed thresholds.
Data feed issues require prompt recalibration; accuracy diminishes beyond 30‑day horizons.
Changes can modify TIR elasticities; periodic re‑estimation preserves edge.
EU–Asia, Trans‑Pacific variants, real‑time API, and contract optimization decision support.
What‑if simulators for fuel shock or capacity‑shock scenarios using stress‑tested ensembles.