A peer-reviewed study published in the open-access journal Frontiers proposes a radial basis function (RBF) neural network to forecast two of container shipping's most punishing variables together — port congestion and freight rates — rather than the single-variable approach that has dominated academic work to date.

The paper models historical vessel-queue and container rate data through a feed-forward network that uses radially symmetric activation functions to map sharp inflection points in the time series. For freight operators who today rely on broker chatter and spreadsheet consensus, that dual-target design attempts to capture a feedback loop single-variable models miss.

Why pair the two forecasts?

When a hub clogs, carriers re-route or bunch sailings, which spikes spot rates, which in turn redirects boxes through already-strained gateways. The transmission is quick and is the source of repeated complaint from shippers on just-in-time retail and automotive lanes.

By forecasting congestion and rates in parallel, the model surfaces simultaneous pressure points a forwarder or BCO planner can act on. Practical consequences look like this:

• For shippers: a one-week improvement in congestion visibility can shift an air-versus-ocean mode choice, where four-figure per-FEU cost gaps between modes are routine.
• For ocean carriers: the same signal feeds extra-loader decisions on sailings that would otherwise leave with half-empty bays.
• For forwarders: the methodology offers a defensible underpinning for rate guidance sold to mid-sized BCOs that cannot subscribe to a full commercial forecasting service.

What makes RBF different from other neural nets?

Linear regression and ARIMA models struggled with the non-linear jumps that defined container markets from 2020 onward, when Asia–US West Coast queues at Los Angeles and Long Beach peaked and the Shanghai Containerized Freight Index climbed by multiples inside twelve months. RBF networks, like LSTM and gradient-boosted alternatives, handle those turns by design.

RBF specifically maps each input through a radial basis — a Gaussian or similar kernel — that produces sharper responses at local points in the data than a standard sigmoid neuron would. That property suits congestion series, where a single week's queue spike matters more than long averages.

The methodology itself is not new outside freight — RBF networks have handled power-load forecasting and currency trading for two decades — but its application to the intertwined port-and-rate problem is rare. Past academic work, the paper argues, has forecast one variable or the other, not both.

What stands between the paper and a commercial tool?

Consistent out-of-sample accuracy across the world's top container gateways. Asia–North Europe, Asia–US West Coast, and intra-Asia pairs each show different congestion signatures, and recent geopolitical shocks — Red Sea diversions via the Cape of Good Hope, drought-driven draft limits at the Panama Canal — have repeatedly broken the assumptions baked into any historical training set.

Industry adoption would likely require retraining on each lane's own series, with continuous recalibration as new shocks distort prior patterns. That labour is feasible for a major carrier's data team; it is expensive for a mid-sized NVOCC operating on narrower margins.

The next test is whether the academic model can survive contact with the data room. If RBF-based forecasting holds up under independent benchmarking against the LSTM and gradient-boosted alternatives that digital freight platforms already deploy, container shipping's analytics stack will gain another input — and the spot-rate guessing game will lose one more excuse.