Written by:
Editorial Team
DSG.AI
This page is a maintained reference for container freight rate forecasting methods, indices, and accuracy benchmarks. It covers the major rate indices, publicly available data sources, statistical and machine learning methods, and published accuracy results. Updated as new results and indices are published.
Last updated: September 2026. See also: Freight Rate Forecasting with Machine Learning: Replacing Analyst Estimates with a Live Model.
Major Container Freight Rate Indices
These are the primary data sources used as inputs and targets in container freight rate forecasting. Understanding what each index measures is the prerequisite for forecasting it.
| Index | Publisher | What it measures | Granularity | Update frequency | Access |
|---|---|---|---|---|---|
| SCFI (Shanghai Containerized Freight Index) | Shanghai Shipping Exchange | Spot rates from Shanghai to major global trade lanes (15 routes) | Per-lane and composite | Weekly (Fridays) | Free |
| WCI (World Container Index) | Drewry | Spot rates on 8 major routes (40ft container) | Per-route and composite | Weekly (Thursdays) | Free composite; paid per-route |
| FBX (Freightos Baltic Index) | Freightos | Spot rates on 12 global trade lanes; updated from live transaction data | Per-lane and global composite | Daily | Free composite; paid per-lane history |
| BDI (Baltic Dry Index) | Baltic Exchange | Dry bulk rates (not container-specific; useful for correlation analysis) | Composite and sub-indices | Daily | Subscription |
| CCFI (China Containerized Freight Index) | Shanghai Shipping Exchange | Contract + spot rates from Chinese ports; broader than SCFI | Composite and per-lane | Weekly | Free |
| IACI (Intra-Asia Container Index) | Shanghai Shipping Exchange | Intra-Asia routes specifically | Per-lane | Weekly | Free |
| NCFI (NINGBO Containerized Freight Index) | Ningbo Shipping Exchange | Spot rates from Ningbo, alternative to SCFI | Composite and per-lane | Weekly | Free |
Which index to target: SCFI and WCI are the most widely cited in academic literature, making them the standard benchmark for method comparisons. FBX is better for operational daily pricing because it updates daily and reflects live transaction data. CCFI captures contract rates that SCFI misses, making it a better predictor of carrier revenue than spot indexes.
Publicly Available Data Sources for Forecasting Models
Beyond the rate indices themselves, freight rate models incorporate macroeconomic, supply, and demand signals.
| Data source | Publisher | Variables available | Granularity | Access |
|---|---|---|---|---|
| Container trade volumes | UNCTAD | Port throughput, trade flows by country pair, TEU volumes | Annual/quarterly | Free |
| Vessel capacity and supply | Alphaliner | Fleet capacity by carrier, orderbook, utilization | Monthly | Free top-100; paid full dataset |
| Port congestion metrics | Flexport Ocean Timeliness Indicator | Days from origin port booking to destination port departure | Weekly | Free |
| Global PMI | S&P Global | Manufacturing PMI by country/region; leading demand indicator | Monthly | Free composite; paid series |
| US retail imports | Global Port Tracker (NRF) | Monthly forecast of container imports at major US ports | Monthly | Free with registration |
| Oil price (Brent) | EIA | Brent crude daily; bunker fuel proxy | Daily | Free |
| AIS vessel position data | MarineTraffic | Vessel position, port calls, idle fleet | Real-time and historical | Freemium; bulk historical paid |
| Panama Canal transit data | Panama Canal Authority | Daily/weekly vessel transits, wait times, cargo volumes | Weekly | Free |
Forecasting Methods: Comparison Table
Published results vary significantly by trade lane, forecasting horizon, and evaluation period. The table below summarizes reported accuracy from peer-reviewed studies and production deployments.
| Method | Typical MAPE (spot rates, 1-4 week horizon) | Typical MAPE (1-3 month horizon) | Implementation complexity | Strengths | Weaknesses |
|---|---|---|---|---|---|
| Moving Average / EWMA | 8–15% | 12–25% | Low | Zero training; stable baseline | Cannot capture structural breaks; lags turning points |
| ARIMA / SARIMA | 6–12% | 10–20% | Low-medium | Captures autocorrelation; interpretable | Stationary data assumption; poor at non-linear relationships |
| VAR (Vector Autoregression) | 5–10% | 9–18% | Medium | Models inter-lane relationships; good for correlated routes | Sensitive to lag selection; instability in high-dimension |
| XGBoost / LightGBM | 4–9% | 7–15% | Medium | Captures non-linear patterns; handles missing data well | Requires feature engineering; no uncertainty quantification native |
| LSTM (Long Short-Term Memory) | 4–8% | 6–13% | High | Captures long-range temporal dependencies | Data-hungry; slow to retrain; harder to explain |
| Transformer (Temporal Fusion Transformer) | 3–7% | 5–12% | High | State-of-art on multivariate time series with external regressors | Computationally expensive; needs large training sets |
| Ensemble (XGBoost + ARIMA + macro features) | 3–6% | 5–10% | Medium-high | Combines statistical regularity with ML pattern recognition | Requires ensemble management; harder to diagnose failures |
| Scenario-based analyst models | 15–30% | 25–50% | Low | Transparent; captures qualitative signals (port strikes, sanctions) | Not systematic; difficult to reproduce; high variance |
Sources: Shi et al. (2023), "Container Freight Rate Forecasting: A Comparative Study," Transportation Research Part E; Fan et al. (2022), "Forecasting Container Shipping Rates Using Machine Learning," Maritime Policy and Management; Duru et al. (2023), "Machine Learning in Freight Rate Forecasting: A Systematic Review," Maritime Economics and Logistics; DSG.AI internal deployment benchmarks.
Interpreting MAPE for freight rates: MAPE is directionally useful but misleading in absolute terms for container freight because rates have high volatility. A 6% MAPE on a $1,000/TEU rate (±$60) is operationally useful. A 6% MAPE on a $500/TEU rate (±$30) may also be within commercial tolerance. The business metric that matters is not MAPE: it is whether the model gets the directional call right (rising vs. falling) in the time window relevant to procurement or pricing decisions.
Published Accuracy Results: Select Studies
| Study | Index targeted | Method | Horizon | Reported MAPE | Notes |
|---|---|---|---|---|---|
| Shi et al. (2023) | CCFI | Temporal Fusion Transformer | 4 weeks | 3.8% | Best result in comparative study; 12 competing methods |
| Fan et al. (2022) | SCFI | XGBoost + macro features | 2 weeks | 5.1% | Feature set: PMI, oil price, port congestion, prior SCFI |
| Chou et al. (2022) | BDI | LSTM | 4 weeks | 4.6% | Dry bulk, not container; included for method comparison |
| Duru et al. (2023) | WCI composite | Ensemble | 1 month | 4.2% | Systematic review; averaged across 5 ensemble variants |
| Yao et al. (2024) | FBX (8 routes) | Transformer + graph attention | 1–8 weeks | 3.1% (1wk), 6.4% (8wk) | Multi-route model capturing inter-lane correlation |
What Indices the Academic Literature Doesn't Cover
Published research skews toward SCFI and BDI because they are free and have long histories. Three gaps matter for production use:
Short-term daily signal. Academic studies use weekly data. Carrier pricing and freight procurement decisions often require daily signals. FBX is the only major index updating daily from live transactions; it is underrepresented in published literature but increasingly relevant for operational pricing.
Contract vs. spot rate divergence. SCFI is a spot index. Most large shipper volumes move on contracts. CCFI includes contracted volumes and diverges significantly from SCFI during market extremes (2020–2022 saw a 300%+ divergence between CCFI spot and contract rates). Forecasting for contract negotiation requires CCFI, not SCFI.
Intra-Asia and emerging routes. Peer-reviewed literature focuses on transpacific and Asia-Europe lanes. Intra-Asia rates (IACI) and emerging trade routes (Middle East, Africa) have minimal published forecasting literature. Academic benchmarks do not transfer.
Production Notes: What the Papers Miss
The gap between academic MAPE and production forecast utility is real. Observations from running freight rate forecasting in production at a tier-1 global container carrier:
- Structural breaks invalidate models trained on prior regimes. The 2020–2022 COVID rate spike and 2023 correction were not captured by any model trained on pre-2020 data. Production models require regime detection and rapid retraining triggers.
- Feature engineering dominates method selection. The difference between a 6% and 4% MAPE is rarely the choice of LSTM vs. XGBoost: it is the quality of the feature set (port congestion leading indicators, forward booking demand, competitor rate signals).
- Forecast uncertainty matters more than point forecast accuracy. Procurement teams and commercial pricing desks need confidence intervals, not point estimates. A model that gives a 5% MAPE with no uncertainty quantification is less useful in production than a 7% MAPE model with calibrated prediction intervals.
For a detailed treatment of how these models are implemented in production at a tier-1 container carrier, see Freight Rate Forecasting with Machine Learning: Replacing Analyst Estimates with a Live Model and Dynamic Pricing for Spot Deals: How a Top 10 Container Line Quotes Against Live Market Data.
This page is maintained by DSG.AI. To suggest additions or corrections, contact us via the maritime page. For commercial maritime AI development, see the maritime AI services overview.
<!-- related-links:start (auto-managed by seo/sync-internal-links.mjs) -->Related
- Vessel ETA Prediction in Production: What the Academic Papers Don't Tell You
- Maritime AI vs. PowerPoint AI: What's Actually in Production in 2026
- Freight Rate Forecasting with Machine Learning: Replacing Analyst Estimates with a Live Model
- Demurrage Prediction: The AI Project That Pays for Itself in One Quarter
- Maritime AI Companies in 2026: Who Ships Production Systems and Who Ships Decks
- Auto-Quotation in Production: From Customer Email to Formatted Quote in Minutes
- Empty Container Forecasting: How We Extended Planning Horizon from 1 Week to 9 Weeks
- FuelEU Maritime Compliance in Year One: What Shipping AI Teams Need to Know
