wenmin-wu/timeseries-gradient-event-boundary-detection

Gradient Event Boundary Detection

Overview

For events with gradual onset and offset (transits, dips, ramps), locate boundaries by computing the first derivative and finding its extrema. The steepest descent marks event start (ingress); steepest ascent marks event end (egress). Works without threshold tuning.

Quick Start

import numpy as np

def detect_event_boundaries(signal, search_start=None, search_end=None):
    """Find event ingress/egress via gradient extrema.
    
    Args:
        signal: 1D array of the time series
        search_start/end: optional tuple (lo, hi) to constrain search region
    Returns:
        (ingress_idx, egress_idx)
    """
    event_min = np.argmin(signal)
    # Search for steepest descent before minimum
    pre = signal[:event_min]
    grad_pre = np.gradient(pre)
    ingress = np.argmin(grad_pre)  # most negative gradient
    # Search for steepest ascent after minimum
    post = signal[event_min:]
    grad_post = np.gradient(post)
    egress = np.argmax(grad_post) + event_min
    return ingress, egress

Key Decisions

• Split at minimum: separates ingress search (before) from egress search (after)
• No threshold needed: uses argmin/argmax of gradient — self-calibrating
• Smooth first: apply Savitzky-Golay or rolling mean before gradient if signal is noisy
• Generalizable: works for any dip/peak event — not just astronomical transits

References

• Source: ariel-only-correlation
• Competition: NeurIPS - Ariel Data Challenge 2024