Data Downloads

Raw prediction records, matched outcomes, and performance datasets for independent analysis

Terms of Use

All data published here may be used freely for non-commercial academic, journalistic, and independent research purposes. Attribution to Seismikon and the evaluation protocol version is requested when results derived from this data are published. Commercial use requires a separate licence - contact commercial@seismikon.com.

Real earthquake data is sourced from the USGS FDSN catalogue and is subject to USGS terms. Our prediction records are provided under the terms described above.

Prospective Predictions

The full set of public prospective predictions issued by Seismikon. Each row represents one forward-issued forecast - a specific latitude, longitude, magnitude, and target time - produced before the forecast window opened.

Fields included:

  • event_uid - unique prediction identifier
  • run_tag - UTC timestamp when the model run was completed (issued time)
  • event_time - UTC timestamp of the predicted earthquake target time
  • latitude, longitude - predicted epicentre (WGS-84 decimal degrees)
  • max_magnitude - predicted magnitude
  • depth_km - predicted focal depth in kilometres
  • eps_empirical - Earthquake Potential Score at time of issue (0–100)
  • grading - outcome label (TP, FP, or null if not yet evaluated)
  • is_current - whether the record was in the active forecast set
  • matched_real_id - USGS FDSN event ID of the matched earthquake (TPs only)
  • matched_real_lat, matched_real_lon - matched earthquake epicentre
  • matched_real_mag - matched earthquake reported magnitude
  • matched_real_depth_km - matched earthquake depth in km
  • matched_real_time_utc - matched earthquake UTC origin time
  • matched_real_dist_km - Haversine distance between prediction and matched event
  • matched_real_region - descriptive region name of the matched earthquake from USGS
⬇ Download CSV ⬇ Download JSON 🔗 JSON API endpoint

The JSON API endpoint (/api/track-record) also returns the current evaluation thresholds (distance, magnitude, time tolerances) so an independent analyst can rerun the matching logic against any earthquake catalogue.

Current Performance Statistics

The live accuracy endpoint returns the latest computed performance metrics (precision, recall*, F₁*, hit rate) along with the thresholds applied. Metrics are computed client-side on the Track Record page and can also be retrieved directly.

🔗 Accuracy score JSON 🔗 Current active forecasts JSON

Model Version History

Seismikon predictions include model_version and algorithm_version fields when available. The table below summarises known model releases.

Model Version Algorithm Version Deployment Date Notes
No model version records available yet.

Evaluation Protocol Version Downloads

Each protocol version is frozen at publication. The current protocol is always available at /evidence/evaluation-protocol. Prior versions are listed here for completeness.

Version Published Distance (km) Mag delta (±) Depth delta (km) Time delta (h) Min real-event mag
1.0.0 current 2025-01-01 100 1.0 100 36 No minimum

Replicating the Track Record Analysis

To independently replicate the track record scoring:

  1. Download the predictions CSV or fetch from /api/track-record.
  2. Fetch real earthquake data from the USGS FDSN API for the same time period and bounding box (the /api/track-record response includes the recommended bounding box).
  3. Apply the match rules from the Evaluation Protocol (version 1.0.0): for each prediction–event pair, all four tolerances must be satisfied simultaneously - spatial (≤ distance_km), magnitude (±magnitude_delta), time (±time_delta_hours), and depth (±depth_delta_km). Build the bipartite adjacency graph from passing pairs, then run maximum bipartite matching (default: Hopcroft-Karp one-to-one; or greedy if multi-match is enabled).
  4. Count TPs, FPs, and FN* as defined in Protocol Sections 8 and 11.
  5. Compute metrics per Section 13.

The browser JavaScript on the Track Record page performs exactly this computation; the source is available at /static/js/track_record_metrics.js.