> ## Documentation Index
> Fetch the complete documentation index at: https://nixtlaverse.nixtla.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Aggregators

> Online aggregation of expert predictions, used by AgACI

### `OnlineAggregator`

Bases: <code>[ArrayModel](#conformalforecast._model.ArrayModel)</code>, <code>[ABC](#abc.ABC)</code>

Online expert aggregator for AgACI and similar methods.

Maintains per-expert cumulative losses and exposes normalized weights.
Subclasses differ in how they convert losses to weights.

The panel shape (`n_experts`, `n_series`, `horizon`) is unknown at
construction time -- it is only known once the caller sees the
calibration arrays -- so it is bound via :meth:`initialize`, not
`__init__`. Call :meth:`initialize` once before :meth:`weights` or
:meth:`update`.

#### `OnlineAggregator.initialize`

```python theme={null}
initialize(n_experts, n_series, horizon)
```

Bind the panel shape and allocate `cumulative_losses_`.

## Parameters

n\_experts : int
Number of experts being aggregated.
n\_series : int
Number of series in the panel.
horizon : int
Forecast horizon length.

## Raises

ValueError
If any argument is \< 1.

#### `OnlineAggregator.update`

```python theme={null}
update(losses)
```

Add per-expert losses to cumulative losses.

## Parameters

losses : NDArray, shape (n\_experts, n\_series, horizon)

## Raises

CalibrationError
If called before :meth:`initialize`.
ValueError
If `losses` does not have the expected shape.

#### `OnlineAggregator.weights`

```python theme={null}
weights()
```

Return current normalized weights.

## Returns

NDArray, shape (n\_experts, n\_series, horizon)
Sum along axis 0 equals 1 for every (series, horizon) cell.

## Raises

CalibrationError
If called before :meth:`initialize`.

### `EWA`

Bases: <code>[OnlineAggregator](#conformalforecast.aggregators.base.OnlineAggregator)</code>

Exponentially Weighted Average aggregator (Vovk 1990).

Weights are proportional to `exp(-eta * cumulative_loss)`. Larger `eta`
converges faster to the best expert but is noisier.

Call :meth:`initialize` with the panel shape before :meth:`weights` or
:meth:`update`.

#### `EWA.initialize`

```python theme={null}
initialize(n_experts, n_series, horizon)
```

Bind the panel shape and allocate `cumulative_losses_`.

## Parameters

n\_experts : int
Number of experts being aggregated.
n\_series : int
Number of series in the panel.
horizon : int
Forecast horizon length.

## Raises

ValueError
If any argument is \< 1.

#### `EWA.update`

```python theme={null}
update(losses)
```

Add per-expert losses to cumulative losses.

## Parameters

losses : NDArray, shape (n\_experts, n\_series, horizon)

## Raises

CalibrationError
If called before :meth:`initialize`.
ValueError
If `losses` does not have the expected shape.

#### `EWA.weights`

```python theme={null}
weights()
```

Return current softmax weights over experts.

## Returns

NDArray, shape (n\_experts, n\_series, horizon)
Numerically stable softmax along axis 0. Sums to 1 along axis 0.

## Raises

CalibrationError
If called before :meth:`initialize`.


## Related topics

- [Methods](/conformalforecast/methods.html.md)
- [Diagnostics](/conformalforecast/diagnostics.html.md)
- [Aggregated adaptive conformal inference](/conformalforecast/docs/methods/agaci.html.md)
- [Architecture](/conformalforecast/docs/concepts/architecture.html.md)