Calculator
Use raw mode to derive each z-score from value, mean, and standard deviation. Use direct mode when your upstream pipeline already produced z-scores.
Example data table
| Feature | Raw Value | Mean | Std Dev | Weight | Z | Weighted Contribution |
|---|---|---|---|---|---|---|
| Prediction Confidence Drift | 0.7800 | 0.5000 | 0.2000 | 0.4000 | 1.4000 | 0.5600 |
| Error Rate Shift | 0.1200 | 0.1000 | 0.0500 | 0.3500 | 0.4000 | 0.1400 |
| Latency Variance Signal | 1.4000 | 1.0000 | 0.2500 | 0.2500 | 1.6000 | 0.4000 |
| Normalized Z Total | 1.1000 | |||||
This example uses weights summing to 1.0000, so normalized Z Total equals the weighted sum.
Formula used
zi = (xi - μi) / σi
ci = wi × zi
Ztotal = Σ(wi × zi) / Σ(wi)
Ztotal = Σ(wi × zi)
Percentile = Φ(Ztotal) × 100Two tailed p = 2 × (1 - Φ(|Ztotal|))
In machine learning workflows, this helps combine standardized feature deviations into one monitoring signal. Positive totals suggest upward deviation, while negative totals suggest downward deviation.
How to use this calculator
- Select Raw values to z-scores or Direct z-scores.
- Enter one or more feature rows with a positive weight.
- In raw mode, provide raw value, baseline mean, and standard deviation.
- In direct mode, provide the z-score from your preprocessing pipeline.
- Choose whether to normalize by total weight.
- Set your absolute threshold for anomaly decisions.
- Click Calculate Z Total to view the result above the form.
- Use the CSV and PDF buttons to export the report.
FAQs
1. What does Z Total represent?
Z Total is one combined standardized score. It summarizes several feature deviations into a single number, making model monitoring, anomaly screening, and comparison easier.
2. When should I use raw mode?
Use raw mode when you have the observed value, expected mean, and standard deviation. The calculator converts each feature into a z-score before aggregation.
3. When should I use direct mode?
Use direct mode when another preprocessing step already standardized your data. This is common in pipelines that store rolling z-scores for live monitoring.
4. Why would I normalize by weight?
Normalization divides the weighted sum by total weight. It keeps the overall score comparable when your weights do not sum to one.
5. What threshold should I choose?
A threshold near 2 is often used for moderate alerts, while 3 is stricter. Your best value depends on how sensitive your monitoring policy should be.
6. Can negative Z Total values be important?
Yes. A negative total means the combined signal is below baseline. Downward drift can matter just as much as upward drift in production systems.
7. What does the percentile mean here?
The percentile estimates where the Z Total sits on the standard normal curve. Higher percentiles indicate more extreme positive movement.
8. What if my standard deviation is zero?
A zero standard deviation makes the z-score undefined. You should revise the baseline, use a stable rolling window, or exclude that feature.