There is a little-known parable in the transportation field: people will generally only remember worst-case traffic conditions.
Think of the gridlocked arterial at 5pm on a Friday before a long weekend, the city blocks that ground to a halt due to an unanticipated closure, or the hours-long crawl on a highway where a major incident disrupted flow.
These are the moments when transportation professionals crave insights to help make informed decisions about how to mitigate these conditions and evaluate whether those mitigations had the desired effect. Here is where probe data shines – at INRIX we can use our massive data-lake to provide a cohesive picture of network operations at scale in real-time and for historic analysis.
These are also the moments when data quality problems are hardest to detect, and when silent model changes do the most damage.
Too many data providers ship updates to their speed models without documentation, without advance notice, and without giving customers any way to evaluate what changed. Baselines shift silently. Analytics drift. And customers are left wondering whether a trend they’re seeing in their data reflects reality or a quiet update they were never told about. We think that’s unacceptable.
When I took over the traffic portfolio at INRIX in early 2025, I made it my mission to ensure INRIX has the best-in-class accuracy metrics. And as a former public sector employee, I knew we need to thoroughly document and communicate any changes well-in-advance with our customer base, both to build trust and allow ample time for analysis.
This approach reflects a deliberate strategic commitment: to be not only the most accurate traffic data provider in the world, but the most transparent one too.
Accuracy at the Extremes
There are well-known and documented deficiencies in the probe data industry at the critical moments detailed in the opening of this blog. They are best summed up in this peer-reviewed study on congestion detection precision from Iowa State.
The researchers found that probe data was meaningfully less precise at detecting non-recurring congestion — the unexpected, incident-driven slowdowns that matter most — with a true detection precision of just 51% when all segment types were included, compared to over 81% for predictable, recurring congestion.
A less-studied but still impactful accuracy issue relates to speeds at the other end. Inaccurate free-flow speeds on high-speed roads create a systematic negative bias that quietly distorts outcomes. The downstream consequences of this bias are documented in a peer-reviewed case study on I-95 Express Lanes, which found that systematic traffic data errors as modest as 5% translated into measurable toll revenue losses, losses that compounded significantly as the error grew.
And while the academic literature on probe data accuracy has historically focused on congestion detection, a 2019 study on probe vehicle trajectories independently confirmed what we observed in our own data: that in high-speed free-flow conditions, probe data consistently skews slower than ground-truth sensor readings — a known industry limitation that, until now, has received far less attention than it deserves.
I knew with a little data science and a lot of hard work, our team could design improvements that deliver more accurate speeds during these tricky conditions on both ends of the spectrum.
A New, and Transparent, Philosophy
As we began planning the algorithmic work that would become 2026.B, we had a broader strategic conversation about what it means to be a trusted data provider. Accuracy improvements are table stakes. But how we communicate changes, and how much access we give customers to evaluate those changes themselves — that’s where we saw an opportunity to differentiate.
We looked at our Map Update process, which runs on a predictable cadence and gives customers advance notice and clear migration windows. It works well. Customers know when changes are coming, they can plan integrations accordingly, and they can test before they commit. We asked ourselves: why don’t we do the same thing for speeds?
The answer is our new “Speeds Update” — or SU — framework. Starting this month, INRIX will publish speed updates on a twice-annual cadence: a beginning-of-year (A) window and a mid-year (B) window. Each update comes with details explaining what changed, why, and what the measured impact is. And critically, any update that affects baseline speeds will include a beta environment where customers can preview the new model running in parallel with production before we flip the switch.
We hosted two town halls to walk existing users through these changes and how to access the beta environment. We distributed a 101-page (phew!) whitepaper to over 10,000 active users. We discussed with each customer during our monthly one-on-ones.
We’re not just telling customers “we got better.” We’re inviting them to verify that for themselves, on their own networks, with their own data. That’s a fundamentally different relationship than the industry has typically offered.
This isn’t just good service. It’s a reflection of how seriously we take the downstream consequences of our data. If you’re building routing logic, congestion pricing models, or safety analytics on top of INRIX speeds, you deserve to know what changed, when, and by how much — before it affects your product.
We believe this should be the industry standard for any probe data supplier, and customers should expect this level of support and communication.
What We Actually Changed: The 2026.B Quality Enhancements
So here, in full, is exactly what we changed and why.
The algorithmic improvements in 2026.B are targeted. We didn’t try to rearchitect the entire INRIX AI Traffic Engine — we went after three specific, well-documented traffic states that customers have raised with us repeatedly.
Trapped traffic detection got a meaningful upgrade. We conducted more in-depth empirical analysis to distinguish brief, ordinary stops from genuine multi-hour entrapment events. We updated the inventory of road segments where trapped-traffic conditions are known to occur, and we introduced new change detection logic that can maintain situational intelligence even when probe data density is low during a standstill. The result is faster and more persistent standstill flagging, and substantially more realistic ETAs during the events that matter most.
Urban stop-and-go accuracy was addressed through a package of model changes focused on state-aware estimation. We reweighted how recent slow-speed probe observations are incorporated into the model, tuned sensitivity to slow probes, and retrained the model to balance ETA accuracy alongside raw speed variance minimization. The practical upshot is reduced error at 5-to-15-minute planning horizons on urban road classes, fewer oscillations during signal cycles, and more stable ETAs during active queueing. The metric we track — Symmetric Mean Absolute Percentage Error (SMAPE) — shows meaningful improvement across the slow-speed bands where this work was focused.
High-speed accuracy was improved by replacing legacy conservative speed caps with a dynamic GPS input speed cutoff filter that is sensitive to both observed and historical conditions by geography. We also broadened the machine learning training set to include higher-speed probe data that previously would have been filtered out, and tuned outlier removal to better distinguish genuinely fast-moving vehicles from GPS artifacts. The net effect is closer alignment to real observed free-flow speeds on major highways, with reduced negative bias — without meaningfully increasing false positives.
Try it Yourself: The Beta Environment
The 2026.B beta environment is live as of June 3, 2026. Customers didn’t need to wait until the production roll-out to start evaluating what’s changing on their networks.
API access is available via a parallel beta endpoint. No schema changes are required — requests and responses are identical in what is already in production. SaaS access through Mission Control and Roadway Analytics is also available. In both platforms, we’ve added a new “Live / Non-Live” feed toggle in the interface, so users can inspect real-time network conditions of both feeds and run historic studies without any API work at all.
For customers who want to run a structured comparison, our white paper includes a sample experiment design: define a geographic area of interest (we recommend no more than 500 segment IDs for the beta environment), pull parallel data from both the production and beta feeds over a consistent time window, and compare speed metrics broken out by functional road class and speed band. That structure mirrors the analysis we ran across 21 countries and tens of millions of data points to validate 2026.B before release.
The 2026.A feed will remain available alongside 2026.B through July 29, 2026, giving customers a full window to validate the transition, annotate any trend breaks in their own analytics, and prepare for the deprecation of the 2026.A endpoint.
Stay Tuned…
In this 2026.B Speeds Update, we didn’t just target quality enhancements, we included a new feature set called Special Lanes. We wanted to bring more granular intelligence to the conditions when two adjacent lanes or lane-types on the same highway are moving at completely different speeds. This has been invisible to data providers until now. In Part 2 of this blog post, I will do a deep dive into the Special Lane typologies, and how this technology enables new use cases.
In the meantime, to request beta access, explore the documentation, or discuss how 2026.B affects your specific integration, contact support@inrix.com or reach out to your INRIX account manager. We built the beta environment because we believe the best quality guarantees aren’t promises — they’re opportunities to look for yourself.