Research & evidence

Evidence first. Claims second.

Before asking partners for operational data, Captrixa tested what official public data can — and cannot — say about short-horizon freight corridor pressure in the Nordics and Baltics.

Approach

How we evaluate.

  • Pre-registered where it matters

    Metrics, baselines and verdict rules are fixed before results are seen.

  • Chronological, no leakage

    Every forecast uses only information available at the time it would have been made.

  • Naive baselines beside every model

    Seasonal and weekday-hour baselines are always reported. A model has to beat them to count.

  • Negative results are recorded

    Where a source shows no usable signal, we say so and do not build on it.

Evidence classes

Four labels that never blur.

Every number in Captrixa carries exactly one of these classes. Modelled opportunity is never presented as savings, and public data is never presented as partner capacity.

  1. RESEARCH

    Research

    Official public and research datasets, used as context.

    StatusFinland, Norway, Estonia and Latvia sources evaluated.

  2. MODELED

    Modelled

    Forecasts, simulations and counterfactuals produced by Captrixa models.

    StatusOut-of-sample 24 h / 48 h forecast evaluations on public data.

  3. OBSERVED

    Observed

    Operational and economic data provided by an approved partner.

    StatusNext stage. Requires design-partner historical data.

  4. LIVE

    Live

    Outcomes of pilot actions that were actually executed.

    StatusNone yet. Only after mandatory shadow mode.

Geographic scope

What the public-data research shows.

Results below are MODELED evidence computed over RESEARCH inputs. Technical details are available in each section.

Norway

Out-of-sample improvement demonstrated

RESEARCHMODELED

Statens vegvesen Trafikkdata — official hourly, per-direction counts (NLOD 2.0)

Freight-relevant road points on Oslo logistics approaches and Sweden-border corridors. In a pre-registered, out-of-sample evaluation, a transparent calendar-aware profile model beat the strongest simple seasonal baseline at both 24 h and 48 h.

Technical details and caveats
  • About 45,600 scored targets per horizon; long-vehicle counts per hour.
  • Mean absolute error skill versus the strongest simple baseline: 22.7% at 24 h (95% CI 18.1–27.5%) and 19.7% at 48 h (95% CI 15.1–24.1%).
  • Excluding public holidays, skill is lower: about 16.9% (24 h) and 14.9% (48 h). Part of the gain is calendar awareness, not a generic “AI uplift”.
  • A probabilistic pressure-hour evaluation showed positive but modest Brier skill (about 3%).

Caveat. The source’s “long vehicle” class (≥ 5.6 m) includes HGVs but also buses and some cars with trailers.

Finland

Modest predictive signal

RESEARCHMODELED

Fintraffic traffic measurement stations (TMS) and Digitraffic Marine port calls

Hourly heavy-vehicle counts on the approaches to the Port of Helsinki (Vuosaari), with a Finland–Estonia ferry interface built from official port-call records. 24 h and 48 h forecasts showed skill over climatology, but modest.

Technical details and caveats
  • Public road and port data are ingested with provenance and revision history; forecasts are evaluated chronologically.
  • A study linking road traffic near Vuosaari to individual ferry departures found no material signal at hourly resolution — reported as a negative result.
  • Port-call data describe ship movements, not freight units or booked capacity.

Caveat. Finnish public data provide corridor context; they do not identify spare or offered truck capacity.

Estonia

Corridor context only

RESEARCH

Transpordiamet traffic counting data — hourly by lane and vehicle class

Counters on key freight roads, including the Via Baltica border crossing. The data are published monthly, so 24–48 h forecasting is not supported by design; the data are used as historical corridor context.

Technical details and caveats
  • Seven permanent counters analysed, with vehicle-class detail.
  • Real-time alternatives require registration and lack vehicle classes and history.

Caveat. Publication lag rules out short-horizon forecasting from this source alone.

Latvia

Corridor context only

RESEARCH

Latvian State Roads — annual traffic volumes and heavy-vehicle shares

Annual traffic and heavy-vehicle shares for the main A-roads, including the Via Baltica sections. Useful for corridor sizing; not for short-horizon forecasting.

Technical details and caveats
  • Official history is annual. The hourly national access-point feed keeps only a few hours and requires registration.

Caveat. Annual data cannot support 24–48 h forecasts.

Sweden and Denmark were assessed. Their verified public traffic datasets are annual or daily, or require separate access for history, so they are not part of the current short-horizon research.

Evidence boundary

Public traffic prediction is not freight capacity prediction.

Official public data is strong evidence for some questions and silent on others. We keep the two apart.

What the public-data research can prove

  • Whether short-horizon corridor pressure contains predictive signal
  • Whether results reproduce across countries and sources
  • How pressure behaves over time — hours, weekdays, holidays
  • Whether a forecasting method beats simple baselines without leakage

What it cannot prove

  • Available commercial truck capacity
  • Partner bookings
  • Carrier acceptance
  • Reserved ferry or trailer space
  • Realised savings
  • Commercial return on investment

What comes next

The next question is commercial, not statistical.

Public corridor data shows whether useful short-horizon signal exists. The next validation step is to determine whether those signals improve real capacity decisions when combined with partner booking, shipment and capacity history.

That next stage tests whether a partner’s own operational data can create measurable decision value against the operator’s current practice: on their route, against their current rules, with the success threshold agreed before the analysis.

Next stage

From public research to partner validation.

  1. Done

    Public-data research

    Short-horizon predictive signal shown on official public data, independently in Norway and Finland.

  2. Next

    Partner historical data

    A bounded period of one partner’s bookings and capacity records on one route, analysed without temporal leakage.

  3. Then

    Shadow mode

    Prospective recommendations recorded beside real decisions — nothing executed.

  4. Only then

    Limited live pilot

    Human approval for every trigger, behind predefined evidence gates.

The next evidence has to come from real operations.

Phase 1 is one route or corridor, a bounded historical study and a private feasibility report — at no cost, with no live integration, no workflow change and no autonomous action.