Skip to main content
Open llms.txtCopy tools

Benchmarks

These reports show where CartoBoost has been measured, what the comparisons mean, and where the evidence is still limited. Start with the result table for the task closest to your own, then inspect its data, split, features, command, runtime, and artifact before transferring the conclusion.

Each report answers the same reader questions:

  • What data was used?
  • What split was used?
  • Which models saw which features?
  • What command produced the artifact?
  • What did the metric table say?
  • Which plots should I inspect?
  • What claim is allowed, and what claim is not allowed?

CartoBoost is evaluated on taxi-shaped structure: pickup and dropoff zones, route distance, periodic hour/day effects, repeated IDs, pickup/dropoff topology, and lane-demand history. The benchmark docs separate real NYC TLC evidence from synthetic mechanism checks and report mixed results when external baselines are lower on a task.

The maintained evidence covers NYC TLC zone and lane tasks, graph forecasting, air-quality interpolation, housing sanity checks, synthetic spatial fields, synthetic graph diffusion, and synthetic geo-causal panels. Real datasets support task-specific quality comparisons; synthetic datasets demonstrate mechanisms and failure modes only.

Report Map

ReportEvidence typeWhat to inspect first
Benchmark MethodologyShared evaluation protocol.Data identity, split rules, metric roster, timing, and reproducibility.
NYC Taxi BenchmarksReal TLC fare, duration, and pickup-demand regression.Current-code CartoBoost versus external baselines, RMSE/MAE/R2/WAPE tables, timing breakdown.
NYC Taxi Path C ClaimsReal TLC tests of geo-temporal behavior.Directional, temporal, known-future, spatial-transfer, and residual-correction results.
Forecasting Tool BenchmarkReal taxi lane demand, synthetic taxi-shaped forecasting, M4 sample, M5 full-roster sample, and M5/M6 full-run protocols.RMSE/WAPE tables, M5/M6 model rosters, run commands, horizon plot, forecast-line plot.
Model Benchmark SuitePublic tabular regression and graph diagnostics.CartoBoost versus external baselines, validation-search selections, full RMSE/MAE/R2/WAPE tables.
Deep Claim BenchmarksSynthetic mechanism checks for deep models.Seven result rows, exact command, and JSON artifact.
Taxi Zone AcceptanceDeterministic taxi-lane feature acceptance.Lane heatmap, hour profile, route midpoint geometry.
Neural Embedding BenchmarkSynthetic repeated-ID/cold-ID diagnostic.Scenario table showing random/tail gains and cold-origin failure.

Current Maintained Artifacts

ArtifactPath
NYC regression JSONdocs/assets/nyc_taxi_benchmarks/results.json
NYC regression JSONL metricsdocs/assets/nyc_taxi_benchmarks/results.jsonl
NYC regression reportdocs/assets/nyc_taxi_benchmarks/results.md
NYC Path C claims JSONdocs/assets/nyc_taxi_benchmarks/path_c_claims.json
NYC Path C claims JSONLdocs/assets/nyc_taxi_benchmarks/path_c_claims.jsonl
NYC Path C claims reportdocs/assets/nyc_taxi_benchmarks/path_c_claims.md
NYC repeated regression JSONdocs/assets/nyc_taxi_benchmarks/repeated_results.json
NYC repeated regression reportdocs/assets/nyc_taxi_benchmarks/repeated_results.md
NYC forecasting JSONdocs/assets/nyc_taxi_benchmarks/forecasting_library_benchmark_real.json
Forecasting suite JSONdocs/assets/nyc_taxi_benchmarks/forecasting_overhaul_committed_suite.json
Forecasting full-roster suite JSONdocs/assets/nyc_taxi_benchmarks/forecasting_overhaul_committed_suite_full_roster.json
M4 forecasting JSONdocs/assets/nyc_taxi_benchmarks/forecasting_overhaul_m4_committed.json
M5 forecasting JSONdocs/assets/nyc_taxi_benchmarks/forecasting_overhaul_m5_committed.json
M6 forecasting JSONdocs/assets/nyc_taxi_benchmarks/forecasting_overhaul_m6_committed.json
Synthetic forecasting suite JSONdocs/assets/nyc_taxi_benchmarks/forecasting_library_suite_synthetic.json
M4 sample suite JSONdocs/assets/nyc_taxi_benchmarks/forecasting_m4_suite_sample.json
M5 full-roster sample JSONdocs/assets/nyc_taxi_benchmarks/forecasting_m5_full_roster_sample.json
M5 full forecasting JSONdocs/assets/nyc_taxi_benchmarks/forecasting_m5_full.json
M6 full forecasting JSONdocs/assets/nyc_taxi_benchmarks/forecasting_m6_full.json
Model diagnostic suite JSONdocs/assets/model_benchmarks_public/results.json
Model diagnostic suite JSONL metricsdocs/assets/model_benchmarks_public/results.jsonl
Model diagnostic suite aggregate JSONdocs/assets/model_benchmarks_public/results_aggregate.json
Model diagnostic suite reportdocs/assets/model_benchmarks_public/results.md
Lane acceptance JSONdocs/assets/lane_level_tests/acceptance_metrics.json
Deep claim benchmark JSONdocs/assets/deep_claim_benchmarks/results.json
Deep claim benchmark JSONLdocs/assets/deep_claim_benchmarks/results.jsonl
Deep claim benchmark reportdocs/assets/deep_claim_benchmarks/results.md

How To Read The Results

A result is usable as benchmark evidence when it names the dataset, command, split, feature policy, models, metrics, and artifact path. It is only a public quality claim when the comparison uses complete required baselines, same rows, comparable feature access, no test-set peeking, equal tuning budget, and uncertainty or repeatability evidence.

Random splits show interpolation. Spatial, grouped, cold-ID, or out-of-time splits are the evidence for deployment risk. Synthetic fixtures are useful for debugging and feature acceptance, not for real-world superiority claims.