Written by:
Editorial Team
DSG.AI
Container freight rate forecasting with machine learning is no longer a research project. A 2025 systematic literature review in Maritime Economics and Logistics catalogued 80+ studies spanning 30 years of ML applications to freight rate prediction. The academic consensus: machine learning models, particularly gradient boosting and LSTM architectures, outperform traditional econometric models on virtually every route and time horizon when the input data is clean. The production consensus, from building this system for a tier-1 global container carrier: the hard part is not the model. It is the data pipeline, the evaluation framework, and the organizational integration that determines whether the forecast changes anything downstream.
Why Traditional Freight Rate Forecasting Fails
Container freight rates are driven by supply (fleet deployment, vessel orders and scrapping, slow steaming decisions) and demand (trade volumes, seasonal patterns, port congestion feedback loops) interacting over multiple time horizons. Traditional forecasting approaches fail for predictable reasons.
Analyst-based estimates anchor on recent history. A rate desk analyst projecting next quarter's Asia-Europe spot rate is running a mental regression against the last two quarters with adjustments for known capacity changes. When conditions shift, the adjustment lag creates systematic overconfidence in trend continuity. The 2021-2022 rate spike and the 2023-2024 normalization both caught analyst estimates consistently behind the actual move.
Index-based forecasting has a publication lag problem. The Shanghai Containerized Freight Index (SCFI), the Freightos Baltic Index (FBX), and World Container Index (WCI) publish weekly. The data reflects rates that were actually traded 5-7 days before publication. A model running on these indices is forecasting future rates from lagged actuals, which is fine for long-horizon planning but insufficient for commercial decision-making that happens on 24-72 hour timescales.
Econometric models assume stationarity that container markets rarely maintain. ARIMA and related time-series models assume the statistical properties of the series are stable over time. Container freight rates are structurally non-stationary: new alliances, canal disruptions, pandemic-era demand shifts, and fleet renewal cycles permanently change the distributional properties of the series. A model calibrated in 2019 is not a useful forecasting tool in 2024.
What the Academic Literature Gets Right
The most recent comparative evaluations confirm several results that hold consistently across routes and time horizons:
Random Forest and gradient boosting outperform ARIMA on most medium-horizon forecasting tasks. The gradient boosting advantage is largest when tabular features are available (fleet capacity data, port congestion indicators, commodity demand proxies) because the models can incorporate them naturally while ARIMA cannot.
LSTM and deep learning models outperform gradient boosting on long-sequence problems where the path dependency in rates matters: the sequence of rate movements over the past 12 months carries information that tabular features summarize poorly. LSTMs capture these trajectories better.
Hybrid architectures outperform single-model approaches when the prediction problem has both a tabular-feature dimension (fleet supply, port state, cargo demand) and a temporal dimension (rate trajectory momentum). The practical implication: use gradient boosting for the feature layer and LSTM or transformer components for the sequence layer.
MAPE is a misleading evaluation metric for this problem. A model that achieves 3% MAPE on Asia-Europe SCFI may be providing almost no decision-useful signal, because the SCFI is itself an average of a range of rates across many ship sizes and cargo types. Percent error on an averaged index obscures prediction quality on the rates that actually matter to your pricing desk.
What Production Requires That the Papers Miss
The published literature optimizes models on historical data with the benefit of hindsight. Production requirements are different.
Real-time data pipelines, not batch datasets. Academic studies train on weekly or monthly index data downloaded as a CSV. A production forecast system ingests live AIS fleet position data, daily port throughput feeds from terminal operating systems, real-time booking data from your own commercial platform, and commodity price indexes on a continuous basis. Building and maintaining these pipelines is 60-70% of the engineering effort and appears in none of the papers.
Carrier-specific signals matter more than market indices. A shipping line's own booking data, equipment availability, and committed capacity on each trade lane carries more predictive signal for that carrier's own pricing than any market index. The carrier has inside information about its own network. A model that uses only public market data is throwing away the most valuable inputs.
The forecast horizon determines the right model and the right inputs. A 48-hour rate forecast for spot pricing decisions uses different features and a different model than a 90-day forecast for capacity planning. Treating freight rate forecasting as a single prediction problem leads to models that are mediocre at both. In production, you build a forecasting stack: different models at different horizons, each optimized for its use case.
Models degrade faster than expected. Container markets shift structurally on 18-24 month cycles due to fleet renewal and alliance reshuffling. A model calibrated before a major alliance restructuring (like the 2024 Ocean Alliance changes) will systematically err until retrained. Monthly retraining is the minimum for short-horizon models; quarterly for longer horizons.
The Architecture: What a Shipping Line Actually Builds
For the tier-1 global container carrier we built this for, the production architecture has four layers:
1. Data ingestion and cleaning. AIS feeds (satellite + terrestrial), terminal data feeds, own booking system exports, commodity indices, vessel schedule databases, and public freight indices. Automated QA: position plausibility checks, duplicate detection, cross-source reconciliation. This layer runs continuously.
2. Feature engineering. Calculated features from the cleaned data: fleet utilization by trade lane, port congestion index (vessels at anchor / average dwell time), seasonal adjustment factors, own booking pace vs. historical, competitive capacity on the lane (from vessel schedule databases and AIS). These features are refreshed on 6-hour and 24-hour cycles depending on source data latency.
3. Model ensemble. Short horizon (0-7 days): gradient boosting on tabular features plus own booking data. Delivered on a 6-hour refresh cycle. Medium horizon (7-30 days): hybrid gradient boosting + LSTM using lagged rate sequences plus fleet position features. Daily refresh. Long horizon (30-90 days): econometric + ML ensemble using fleet capacity pipeline data and commodity demand forecasts. Weekly refresh. Each model outputs a point forecast plus confidence intervals.
4. Commercial integration. Forecasts feed into the rate desk platform as a reference signal alongside analyst judgment. The key integration decision: the model is a decision support tool, not an automated pricing engine, at least in the first deployment. The rate desk reviews model output and flags disagreements. Those flags become training data for model improvement.
| Forecast horizon | Primary model | Key inputs | Commercial use case | Refresh |
|---|---|---|---|---|
| 0-48 hours | Gradient boosting | Own booking data, port queue | Spot pricing decisions | 6-hour |
| 7-30 days | Hybrid GB + LSTM | Fleet position, port congestion, booking pace | Capacity allocation, client negotiations | Daily |
| 30-90 days | Ensemble + econometric | Fleet supply pipeline, commodity demand | Network planning, vessel deployment | Weekly |
What This Delivered
The primary outcome for the carrier was the elimination of the 2-4 hour manual quoting cycle for spot rate inquiries. Before the system, a rate inquiry came in by email, an analyst pulled current index data and competitor information, applied judgment-based adjustments, and returned a quote. Hours. After deployment, the model generates a pricing recommendation in minutes from the same inputs. The quoting cycle compressed from hours to minutes. Commercial teams responding to inquiries in this window convert at materially higher rates than those who respond in hours.
The second-order outcome: consistent pricing logic across the rate desk. Manual analyst estimates vary by who is on shift and what information they have seen recently. Model-based recommendations apply the same feature set and logic to every inquiry. Rate consistency improves, and the model's blind spots become visible when analyst overrides cluster around specific conditions.
A third outcome that took longer to materialize: better 90-day capacity planning. When the long-horizon model's capacity projections are sufficiently reliable, vessel deployment decisions can be made further out, reducing the frequency and cost of late repositioning.
What to Build First
If you are starting from a standing start, the sequencing matters.
Build the data pipeline before the model. A model trained on dirty or lagged data will be confidently wrong in ways that are hard to diagnose once in production. Get clean, reliable feeds for your own booking data and AIS fleet positions before you touch model architecture.
Build the short-horizon model first. It has a faster feedback loop: you know within days whether the 48-hour forecast was right. Long-horizon validation takes months and requires you to trust a model you cannot yet verify.
Integrate into existing commercial workflows before automating decisions. The first deployment of a freight rate forecasting model should be a reference signal that rate desk analysts can confirm or override. This creates training data, builds trust in the model, and surfaces systematic errors before they propagate into autonomous pricing decisions.
The vessel ETA prediction work we described in our ETA prediction production article uses the same AIS data infrastructure that feeds the freight rate model. Once the pipeline exists, the marginal cost of adding additional models against the same data drops significantly. For the full range of commercial AI applications we have built into production for container carriers, see our Maritime AI overview.
