Understand dataset spread using mean absolute deviation instantly. Select mean or median centers for robustness. Download clean results as CSV or PDF, anytime securely.
| Observation | Value | Notes |
|---|---|---|
| 1 | 12 | Lower boundary |
| 2 | 15 | Typical value |
| 3 | 18 | Typical value |
| 4 | 20 | Upper mid |
| 5 | 22 | Higher boundary |
Mean Absolute Deviation (MAD) measures the average absolute distance between data points and a chosen center.
Scaled MAD is MAD × 1.4826, commonly used to compare with standard deviation for normally distributed data.
MAD summarizes typical spread using average absolute distance from a chosen center. Unlike squared-error measures, every unit of deviation contributes linearly, so a single extreme point cannot dominate the metric. In feature engineering, MAD helps flag columns that are nearly constant, overly noisy, or drifting between training and production. It is ideal for quick QA in notebooks.
MAD is expressed in the same units as the data: seconds, rupees, clicks, or temperature. This makes it easy to communicate. For example, a MAD of 2.4 minutes means observations differ from the center by about 2.4 minutes on average. When comparing multiple features, consider rescaling or standardizing first, because larger-valued features naturally show larger MAD. A helpful relative view is MAD divided by the mean or median.
Using the mean as the center aligns MAD with many regression and forecasting workflows, especially when the mean is a meaningful operating target. Using the median improves robustness when distributions are skewed or contain spikes. The calculator reports both mean and median so you can see how asymmetry changes the center and, therefore, the deviations. If the mean and median diverge materially, it often signals skewness or mixed populations.
Optional trimming removes a percentage from each tail before computing statistics. This is useful for telemetry spikes, bot traffic, or sensor glitches. Common starting points are 1% to 5% per tail, but ensure enough observations remain for stable estimates. The page also shows Q1, Q3, and IQR, giving a quick view of central spread. If IQR is small but MAD is large, you may have isolated extremes worth investigating.
For audits and monitoring, reproducibility matters as much as the metric. Exporting the breakdown table preserves each value, its deviation, and its absolute deviation, making reviews faster and simplifying peer checks. The scaled MAD (MAD × 1.4826) provides an approximate standard deviation under normality, which can help when aligning dashboards that already track standard deviation. Use CSV for pipelines and PDF for sharing.
It measures the average absolute distance of values from a chosen center. Because distances are not squared, the result is easy to interpret in the original unit and is less dominated by extreme points.
Use the median when your data is skewed, contains spikes, or has clear outliers. Median-centered MAD stays stable under isolated extremes, which is useful for operational metrics, latency data, and spend distributions.
Trimming removes a percentage from each tail before computing the center and deviations. This can reduce the impact of anomalies, but it also changes the sample size, so document the trim rate for repeatable reporting.
Scaled MAD multiplies MAD by 1.4826 to approximate standard deviation when data is roughly normal. It helps compare robust dispersion to dashboards or models that already use standard deviation as a baseline.
Yes. Paste the error series (actual minus forecast) to summarize typical absolute error around the selected center. If you center on zero, the MAD aligns closely with mean absolute error for many workflows.
MAD works for small samples, but stability improves with more observations. For monitoring, aim for at least 30 points per window. If you trim, ensure the remaining count still represents typical behavior.
Important Note: All the Calculators listed in this site are for educational purpose only and we do not guarentee the accuracy of results. Please do consult with other sources as well.