> For the complete documentation index, see [llms.txt](https://docs.everesteer.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.everesteer.ai/research-guide/ensembling-and-auxiliary-targets.md).

# Ensembling and auxiliary targets

Using the auxiliary targets and the benchmark to build a differentiated signal.

The dataset includes auxiliary targets alongside the primary `target_everest`. Each is a related forward-looking construction on the same underlying data, at one of several horizons. Training separate models on different targets and blending the predictions can lift your AIMC and NCORR.

## The approach

{% stepper %}
{% step %}

### Train one model per target

The platform's hosted trainer fits one target per job. Train a separate model for each target you want to include. Each job is metered; preview cost with `dry_run=True`.

Targets available on the tournament dataset: `target_everest` (its primary payout target), `target_Tougroute`, `target_Meltsene`, and others. The full list is in `eiq_features.json` and `get_dataset_schema`. Another dataset, such as the one an event serves, can publish its own target block under its own names; on any dataset, `primary_target` on the schema response is the column you are graded on and the one your blend must predict.
{% endstep %}

{% step %}

### Blend the predictions

Once you have predictions from each target-specific model, blend them into a single prediction for `target_everest`. A simple average of the predictions is a common starting point. Weighted blends or stacking can improve further.

You still submit a single `target_everest` prediction. The blend is what you submit.
{% endstep %}

{% step %}

### Pick genuinely different targets

Some auxiliary targets are near-duplicates of the primary target and add little ensemble value. On the `train` split, `target_Tougroute` tracks `target_everest` closely (correlation near 0.84). Pick targets that produce meaningfully different predictions.
{% endstep %}

{% step %}

### Feature-neutralise

Reducing exposure to dominant feature groups lifts AIMC and NCORR. The simplest approach: after training your model, measure its correlation with each feature column (the SDK's `feature_exposure` function reports the max absolute Spearman). If a single feature dominates, neutralise the predictions by removing the linear projection onto the feature set.

The public notebook `himalayas/feature_neutralization_&_ensembling_everesteer.ipynb` in the example-scripts repository walks through the full process. The repository is private until go-live; collaborators can access it now. The `get_started` tool works regardless of repository visibility.
{% endstep %}

{% step %}

### Use the benchmark files

Benchmark predictions are served as downloadable parquet files matching each split: `eiq_train_benchmark_models.parquet`, `eiq_validation_benchmark_models.parquet`, `eiq_live_benchmark_models.parquet`. The benchmark column is named `v1_sherpa`. Download the benchmark file for your split to measure your predictions' correlation to the designated benchmark.
{% endstep %}
{% endstepper %}

## Why this works

AIMC rewards predictions that differ from the designated benchmark while being correct. A single model trained only on `target_everest` may converge to a solution that correlates highly with the benchmark. Ensembling across diverse targets produces predictions that are differentiated from the benchmark while still being predictive of the primary target. NCORR rewards signal that survives neutralisation against the core feature set; feature neutralisation during ensembling directly targets this.


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://docs.everesteer.ai/research-guide/ensembling-and-auxiliary-targets.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.