> For the complete documentation index, see [llms.txt](https://docs.seldon.ai/mpm/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.seldon.ai/mpm/supported-metrics/calculating-metrics.md).

# Calculating Metrics

Metrics are computed over a user-specified **time window**.\
The specified time window is then split into **time buckets**, each of which having a duration equal to a user-specified **interval**.

## Time Window

Specify **the time window** for the metrics calculation by providing a **start datetime** and an **end datetime**, using the ISO 8601 format (e.g., `2025-01-01T00:00:00.000Z`).

{% hint style="info" %}
The end datetime is open, meaning it is *exclusive*, and the same applies to the end datetimes within each of the time buckets.
{% endhint %}

## Interval

The **interval** parameter determines the size of each bucket within the time window and must be specified as a duration string. A duration string consists of decimal numbers, optionally including fractions, followed by a unit suffix such as `10s` for 10 seconds, `2h45m` for 2 hours and 45 minutes, or `1.5h` for 1.5 hour. Valid time units include `us` (or `µs`), `ms`, `s`, `m`, `h`.

{% hint style="warning" %}
Although `ns` (nanoseconds) is a valid time unit from the protobuf perspective, it is not meaningful in this context, as PostgreSQL database used for storing predictions and ground truth supports only `us` (microsecond) resolution.
{% endhint %}

Consider the following important points:

* The interval value must not exceed the *duration of the time window*.
* If the interval parameter is not provided, a default value of one bucket matching the entire time window is used.

## Time Bucketing

The time window is divided into buckets of the specified interval size, starting from the time window's start datetime.

If a *non-integer number of intervals* fits within the time window, the last bucket will have the following start and end datetimes:

$$
\left\[ \texttt{time window start datetime} + \texttt{math.floor} \left( \frac{\texttt{time window duration}}{\texttt{interval}}\right) \times \texttt{interval}, \texttt{time window end datetime}\right)
$$

In this case the final bucket will be smaller than the specified interval.

{% hint style="warning" %}
The **maximum number of buckets is 100**; therefore, the interval and the time window should be selected accordingly.
{% endhint %}

A data point is assigned to a bucket if the model's inference time falls within that bucket. Each bucket in the Metrics API response includes an `endTime` field.

{% hint style="info" %}
If there are no data points within the time bucket:

* `-1` will be returned for scalar metrics such as Accuracy, Recall, Precision, Specificity, F1, MSE, MAE, and RMSE.
* An **empty array** will be returned for non-scalar fields, such as the confusion matrix fields.
  {% endhint %}