Analysis · Finance
Doubling Traffic in a Model Is Not a Forecast of Traffic Growth
By AWEI · AI-compiled · Published · Analysis prepared · 1 source · doi.org
Traffic demand doubled inside a computer model, and a forecasting algorithm still produced strong scores. But the doubling was an assumption supplied by researchers, not a prediction about how much traffic would grow. That distinction is central to reading this 2021 study by Rusul Abduljabbar, Hussein Dia and Pei-Wei Tsai.
The team calibrated a simulation of Melbourne’s Eastern Freeway using field measurements from 55 detectors. They then generated traffic counts, speeds and occupancy—the percentage of time vehicles cover a detector—for the morning peak. Forecasting models learned from the first 60% of the simulated sequence and were tested on the remaining 40%, predicting five to 60 minutes ahead.
Their bidirectional long short-term memory model, or BiLSTM, processes input sequences in both directions. For vehicle counts 60 minutes ahead, it scored 93.04%, compared with 68.71% for the unidirectional version. These scores mean 100 minus mean absolute percentage error; they do not mean that 93% of individual forecasts were correct.
The researchers also increased simulated demand by 25%, 50%, 75% and 100%, testing the existing model without retraining. Results varied by outcome. With demand doubled, the 60-minute score was 93.00% for counts but only 81.52% for occupancy. The paper’s concluding claim that all three outcomes exceeded 92% across these scenarios overstates its own table.
This was a forecasting-method evaluation using simulated data, not evidence that deploying the algorithm causes less congestion. Calibration against actual traffic supports the simulation’s relevance, but does not establish performance under real future conditions. The analysis covered one freeway’s weekday congestion, excluding weekends, arterial roads and weather inputs.
For readers assessing market forecasts, the useful lesson is methodological: separate assumed growth from predicted outcomes, identify exactly what an accuracy score measures, and check results by category and forecast horizon. Success within a constructed scenario supports a bounded claim about model performance, not the likelihood of that scenario occurring.
Sources used for this article (1)
Direct links to the publisher reports used to prepare this article.
- Source 1
- Development and evaluation of bidirectional LSTM freeway traffic forecasting models using simulation data Rusul Abduljabbar, Hussein Dia, Pei‐Wei Tsai
