sodermalm · page built 2026-09-19 04:09 UTC

Coverage, last 11 runs

Forecast mode. The dashed line is the 80 % the interval claims; below it the published band is too narrow.

Last night

LAST NIGHT  ·  sodermalm  ·  2026-09-19
=======================================

  STATUS   thin

      Only 0 street(s) reported, below the 5 this model
      needs to say anything about the network as a whole. The night was
      processed and the model learned little from it, deliberately.

Honesty statement

HONESTY STATEMENT  ·  sodermalm  ·  as of 2026-09-19
====================================================

Of the times we published an 80 % interval, how often did the truth fall inside it?

  STREETS WE MEASURE

      estimate mode    80.3%   honest
                      median error 0.097 log  (x1.10)
      forecast mode    68.9%   TOO NARROW by 11 points
                      median error 0.177 log  (x1.19)
      over 5,191 nights.  The gap between the two modes is what knowing the day
      is worth, measured rather than asserted.

      Lately, on the same streets, cell-weighted the same way:

        window          estimate  forecast   nights    cells
        all time           80.3%     68.9%    5,191        -
        last 365 days      90.8%     82.3%      175   56,559
        last 90 days       92.7%     81.3%       88   39,903

      CALENDAR days, not the last N scored nights: the sensor fleet thinned, so
      the last 365 SCORED nights span over three calendar years here and would
      be a misleading thing to call 'the last year'. A window with few nights is
      a weak signal, which is why the counts are beside the figures.

  STREETS WE DO NOT

      *** THIS NUMBER IS FROM AN OLDER RUN. validation_edges.parquet predates
          the model beside it, so the two describe different fits. Re-run with
          validation enabled - `pipeline <area> --until <date> --kfold` - before
          quoting it. ***
      coverage         48.5%   TOO NARROW by 31 points
                      median error 0.704 log  (x2.02)
      over 15 streets whose counts were hidden and then predicted exactly
      as an unmeasured street is. This is the claim the product rests on.

      Worst-covered, which the pooled figure hides:

        3357373856929621821   coverage  1.2%   error x11.14   (168 cells)
        6079980744524669011   coverage 22.4%   error x12.03   (161 cells)
        8954425914444133656   coverage 23.4%   error x3.41   (4280 cells)
        7831589162891272994   coverage 35.5%   error x2.09   (791 cells)
        7122657651002101828   coverage 40.7%   error x1.69   (140 cells)

      A pooled number blends a well-calibrated majority with a few streets that
      are confidently wrong. Both are worth reporting; only one is a tuning
      problem.


HOW MUCH IS REPORTING
---------------------

      21 streets report on a typical recent night (at or above the 5 the day factor needs).

  HOW MUCH OF THE NETWORK IS MEASURED AT ALL

      1,109 of 4,564 streets (24.3 %) have ever been counted.
      The other 3,455 are inferred from the network. A city where
      3 % of streets carry recent measurement is a different product from one where
      30 % do, and a uniform-looking map should not imply uniform confidence.

      Of the 1,109 counted, 43 were counted within 120 days and
      carry the narrow measured-street band. The remaining 1,066 are STALE -
      counted once, long ago - and are banded as inferred streets, because that
      is what they now are. Their level comes from the drift and their neighbours,
      not from their own sensor.

Change alerts

CHANGE ALERTS  ·  sodermalm  ·  as of 2026-09-19
================================================

7 confirmed changes. A change is reported when the network's day factor stays displaced for a run of dates, not when a single day looks unusual.

  2010-03-10  UP     +33.2%
      began        2010-03-10   (when the shift started, not when it was noticed)
      confirmed    2010-05-05   (105 processed dates in the run)
      size         +33.2%  (+0.287 log)   peak statistic 10.0

  2010-08-19  UP     +66.0%
      began        2010-08-19   (when the shift started, not when it was noticed)
      confirmed    2010-10-14   (113 processed dates in the run)
      size         +66.0%  (+0.507 log)   peak statistic 11.4

  2012-05-24  UP     +30.8%
      began        2012-05-24   (when the shift started, not when it was noticed)
      confirmed    2012-06-16   (27 processed dates in the run)
      size         +30.8%  (+0.268 log)   peak statistic 10.2

  2013-11-28  UP     +31.5%
      began        2013-11-28   (when the shift started, not when it was noticed)
      confirmed    2013-12-21   (24 processed dates in the run)
      size         +31.5%  (+0.274 log)   peak statistic 10.1

  2020-03-15  DOWN   -23.5%
      began        2020-03-15   (when the shift started, not when it was noticed)
      confirmed    2020-03-31   (54 processed dates in the run)
      size         -23.5%  (-0.268 log)   peak statistic 10.3

  2021-08-12  UP     +24.6%
      began        2021-08-12   (when the shift started, not when it was noticed)
      confirmed    2021-09-04   (25 processed dates in the run)
      size         +24.6%  (+0.220 log)   peak statistic 10.4

  2022-05-05  UP     +23.7%
      began        2022-05-05   (when the shift started, not when it was noticed)
      confirmed    2022-06-12   (39 processed dates in the run)
      size         +23.7%  (+0.212 log)   peak statistic 10.2

What this does NOT tell you:

  · which streets. This watches the whole network's day factor, so it catches a shift
    that moved everything. A single street changing is a different test (paper §8.3).
  · why. Weather, roadworks, a new bridge and a policy change look identical here.
  · that the model has corrected for it. A confirmed change should trigger a local reset;
    until it does, forecasts for the affected streets stay wrong while every daily output
    continues to look right (paper §8).

Each day: what the sensors said, and what the model published

Predicted flow across the whole network, hour by hour.

The model behind these numbers

Recent runs