Calculator Controls
Formula Used
Mean-shift score
Variance-shift score
CUSUM score
Robust threshold
The combined method joins mean, variance, and trend evidence. Penalties reduce scores before threshold testing. Selected points must also satisfy distance and direction rules.
How to Use This Calculator
Paste observations in their true chronological order. Select the expected change type and comparison scope. Local windows work well for repeated regime changes.
Set minimum segments before adjusting sensitivity. Add smoothing when noise hides sustained changes. Use penalties when the calculator finds excessive boundaries.
Submit the form and inspect both charts. Compare markers with segment statistics and confidence values. Confirm important changes using operational or domain evidence.
Understanding Change-Point Analysis
Change-point analysis finds moments where a sequence changes its behavior. These changes may affect averages, variation, trends, or observation patterns. Detecting them helps models respect different operating regimes.
A stable series usually fluctuates around consistent statistical properties. A change creates evidence that one segment differs meaningfully. The calculator converts that evidence into ranked candidate positions.
How Scoring Methods Work
CUSUM accumulates signed deviations from a chosen reference level. Persistent movement pushes the cumulative score beyond a decision threshold. This method reacts quickly to sustained mean shifts.
Peak selection ranks the strongest remaining high-scoring candidate boundaries. Each candidate compares left and right segment statistics carefully. Minimum distance prevents nearby duplicate or statistically trivial divisions.
Advanced Detection Controls
Mean-shift scoring compares averages before and after each candidate. Variance-shift scoring compares squared dispersion across neighboring segments. Combined scoring highlights changes affecting both location and spread.
Robust mode reduces extreme-value influence using median-based scaling. Smoothing can reduce noise before candidate scores are calculated. Penalties discourage excessive change points in volatile sequences.
Choosing Reliable Settings
Strong results need enough observations around every possible boundary. Small segments can produce impressive but unreliable score differences. Larger windows improve stability while reducing timing precision.
Confidence values describe relative evidence, not guaranteed causal certainty. Domain events should confirm whether detected breaks make practical sense. Analysts should review nearby values and segment summaries together.
Running the Analysis
Paste ordered values using commas, spaces, or separate lines. Choose a detection method matching the expected change type. Then set sensitivity, smoothing, penalties, and segment limits.
Submit the form to calculate candidates and segmented statistics. The result appears above the controls for immediate review. Hover over plotted markers to inspect exact values and scores.
Practical Machine Learning Uses
Detected shifts can reveal sensor faults, traffic changes, or demand transitions. They also expose training data drift and production behavior changes. Segments may support separate models, alerts, or retraining schedules.
A single point should rarely trigger an automatic operational decision. Combine statistical evidence with context, costs, and monitoring history. Repeated analysis builds stronger baselines for future shift detection.
Interpreting Advanced Outputs
Segment tables reveal local means, deviations, lengths, and ranges. These summaries show how strongly each regime differs. They also expose short regions needing cautious review.
Score normalization makes candidates easier to compare across methods. Higher rankings indicate stronger separation under selected assumptions. Different settings may reorder candidates significantly across changing datasets.
Exported results can support reports, audits, and experiments. Save chosen parameters beside every detected boundary. Reproducible settings make later comparisons more trustworthy overall.
Frequently Asked Questions
What is a change point?
A change point is a boundary where statistical behavior shifts. The mean, variance, trend, or several properties may change. It separates observations into regimes requiring different interpretations.
Which detection method should I choose?
Use mean shift for level changes and variance shift for volatility. Trend shift targets slope changes. CUSUM finds persistent deviations. Combined evidence is useful when the change type remains uncertain.
Why were no change points detected?
The threshold may be too strict for available evidence. Reduce sensitivity, lower penalties, or expand local windows. Also verify that each regime contains enough observations.
Can smoothing hide real changes?
Yes. Large smoothing windows can blur abrupt boundaries and delay detected positions. Start without smoothing, then increase it gradually when noise dominates the series.
What does the confidence value mean?
Confidence ranks score strength above the selected threshold. It is a relative indicator created by this calculator. It does not prove causation or guarantee a true structural break.
How large should each segment be?
Segments should contain enough observations for stable estimates. Five to thirty points may work for many examples. Noisy or seasonal data often needs larger segments.
Can this detect multiple changes?
Yes. Set the maximum change count above one. Minimum distance and peak suppression prevent nearby candidates from representing the same underlying boundary.
Should I normalize the data?
Normalization helps compare series measured on different scales. Robust scaling is useful with extreme values. It usually preserves positions while changing score magnitudes.
Can results support production monitoring?
Yes, but detected changes should trigger review rather than automatic conclusions. Combine boundaries with business events, model metrics, alerts, and retraining policies before acting.