Prediction Track Record

California public demonstration - forward-issued predictions vs. observed seismic events

Prospective All results on this page are from live, forward-issued predictions evaluated after the fact against real seismic events - not retrospective back-testing. California is the region we have chosen for our public demonstration because it has a high-quality seismic catalogue. It is not a claim that the system has only been designed for California; the same methodology and evaluation protocol apply to any covered region.

Event Time (UTC)	Lat / Lon	Mag	Depth	Event ID	Status	Matched Prediction
Loading…

Predicted Time	Lat / Lon	Mag	Depth	UID	Result
Loading…

Prediction Outcomes Prospective

Precision ⓘ

TP / (TP + FP)

Recall ⓘ

TP / (TP + FN)

F1 Score ⓘ

Harmonic mean

Performance Breakdown Based on the selected time window above

Recall by Magnitude Band ⓘ

Epicentral Distance Error ⓘ

Magnitude Error (ΔM) ⓘ

Lead Time Distribution ⓘ

Time (UTC)	Location	Magnitude	Depth	EPS	Status

Precision

TP ÷ (TP + FP)

Of all earthquakes we predicted, what fraction actually happened? High precision = few false alarms.

Recall

TP ÷ (TP + FN)

Of all real earthquakes in the window, what fraction did we predict? High recall = few missed events.

F1 Score

2 × (P × R) ÷ (P + R)

The harmonic mean of Precision and Recall. Penalises being good at only one of the two - a model that predicts everything scores high recall but low precision and gets a poor F1.

Accuracy

(TP + TN) ÷ (TP + FP + TN + FN)

The fraction of all decisions that were correct. In earthquake forecasting, True Negatives (TN) are undefined - we can’t enumerate every place-time where nothing was predicted and nothing happened - so accuracy in the traditional sense is not computable here.

% Within Tolerance

TP ÷ total predictions

Also called Precision here. Of all predictions issued, what percentage fell within the location, time, magnitude and depth tolerances of a real earthquake? This is not the same as accuracy: it only measures whether predicted events occurred, not whether all real events were predicted.

TP / FP / FN

Confusion matrix cells

TP (True Positive): a prediction that was matched to a real earthquake within all four tolerances. FP (False Positive): a prediction with no matching real event — a false alarm. FN (False Negative): a real earthquake that no prediction covered — a missed event.

EPS (Empirical Probability Score)

Model-specific confidence score

An internal confidence score produced by the model for each prediction. Higher values indicate greater model certainty. Not directly comparable across model versions.

Lead Time

quake time − issued time

For matched predictions, the time between when the forecast was published and when the earthquake occurred. A longer lead time means more advance warning.

Prospective vs. retrospective: All results shown here are from live, forward-issued predictions evaluated after the fact - not back-tested on historical data. The scoring window ends 48 hours before the current time to allow for seismograph confirmation delays.

Prediction Track Record

Debug Panel

Prediction Outcomes Prospective

Model Performance Stats

Before You Proceed

Prediction Track Record

Debug Panel

Prediction Outcomes Prospective