AgiMo · C6

Metrics & evaluation

Raw errors, baselines and the evidence behind a comparison.

A metric for each type of answer

TargetPrimary metricInterpretation
Traffic volumemean |ln(prediction / reference)|Lower is better. Exponentiating gives a typical error factor: 1× is exact. Also report MAE in vehicles/day.
Heavy-vehicle shareMAE (pp)Lower is better. Predicting 8% instead of 5% gives an error of 3 percentage points.
Choice / platform categoryAccuracy + balanced accuracyHigher is better. Show class frequencies; accuracy alone can favour the majority category.
Exclusive mode probabilitiesBrier: mean Σ(p − y)²Lower is better, range 0–2. Probabilities must sum to 1. Secondary: log loss.
Independent mode-use probabilitiesBrier: mean over people and labelsLower is better, range 0–1. Probabilities do not need to sum to 1.
Policy scenariosSeparate judge rubricReport each axis and human agreement. No factual-accuracy score or cross-task average.

These metrics define the new protocol. Historical score formulas and dates remain unchanged. Partial numeric/probability runs receive no primary score; invalid classification answers count as wrong.

Metric definitions: scikit-learn documentation · Research on log accuracy ratios

Baseline results on held-out items

These are computed statistical baselines, not language-model results. Fit uses development items only. Lower loss is better; accuracy is higher-is-better. Brackets show 95% bootstrap intervals.

PackDev / TestPrimary metricConstant baselineFitted baseline
Dresden heavy-vehicle share · extension96 / 22MAE (percentage points)1.273[0.909, 1.636]Not fitted
Dresden stop platforms · transport mode186 / 54accuracy0.574Balanced accuracy: 0.200[0.444, 0.704]Not fitted
Dresden traffic volume · extension96 / 22mean absolute log ratio0.556×1.74[0.391, 0.735]Not fitted
Intercity choice · probability protocol v2168 / 42multiclass Brier (sum over classes)0.732[0.707, 0.762]0.489Regularised conditional multinomial logit[0.390, 0.593]
Klang Valley mode use · probability protocol v2152 / 48Brier (mean over labels)0.178[0.154, 0.203]0.121Three independent regularised logistic regressions[0.086, 0.157]

Held-out sets are small and public. Volume and heavy-vehicle tasks share streets. Mode-choice results describe the selected survey samples, not Dresden or population mode shares.

Existing model answers in raw metrics

Recomputed diagnostics from saved answers; the original result tables are preserved. Intervals reflect variation across questions or streets, not model repeats. Numerical differences alone do not establish a ranking.

Traffic volume · historical 10 segments
ModelRaw result95% intervalValid / asked
deepseek0.712mean absolute log ratio×2.040.319 – 1.19110 / 10
gemini0.346mean absolute log ratio×1.410.147 – 0.60210 / 10
gemma0.735mean absolute log ratio×2.080.207 – 1.37910 / 10
glm0.751mean absolute log ratio×2.120.304 – 1.34210 / 10
glm530.711mean absolute log ratio×2.040.305 – 1.24410 / 10
gpt50.796mean absolute log ratio×2.220.336 – 1.35710 / 10
gptoss0.837mean absolute log ratio×2.310.348 – 1.44810 / 10
kimi0.781mean absolute log ratio×2.180.294 – 1.37210 / 10
llama0.696mean absolute log ratio×2.000.278 – 1.19310 / 10
minimax0.717mean absolute log ratio×2.050.329 – 1.25710 / 10
minimax30.680mean absolute log ratio×1.970.220 – 1.25610 / 10
qwen0.693mean absolute log ratio×2.000.210 – 1.30010 / 10
Heavy-vehicle share · historical 15 segments
ModelRaw result95% intervalValid / asked
deepseek3.767MAE (percentage points)1.833 – 6.40015 / 15
gemma7.600MAE (percentage points)5.833 – 9.50015 / 15
glm4.600MAE (percentage points)2.267 – 7.20015 / 15
glm534.333MAE (percentage points)2.333 – 7.26715 / 15
gptoss6.833MAE (percentage points)4.900 – 8.83315 / 15
llama7.067MAE (percentage points)4.400 – 10.20015 / 15
minimax34.067MAE (percentage points)2.133 – 6.93315 / 15
qwen3.600MAE (percentage points)1.567 – 6.53315 / 15
Traffic choice · historical 12 questions
ModelRaw result95% intervalValid / asked
gpt50.583accuracy0.250 – 0.83312 / 12
gemini0.500accuracy0.250 – 0.75012 / 12
kimi0.333accuracy0.083 – 0.58312 / 12
Traffic choice · historical 60 questions
ModelRaw result95% intervalValid / asked
gemma0.417accuracy0.300 – 0.53360 / 60
llama0.317accuracy0.200 – 0.43360 / 60
glm0.400accuracy0.283 – 0.53360 / 60
minimax0.400accuracy0.267 – 0.51760 / 60
gptoss0.317accuracy0.200 – 0.43360 / 60