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

# Understanding forecasts

> Choose the explanation that matches the question you want to answer.

TimeGPT can help you understand both a forecast and the historical information
behind it. Start with the question you are trying to answer.

## What would you like to understand?

> [card] [Why is this forecast high or low?](https://nixtla.io/docs/forecasting/exogenous-variables/interpretability_with_shap.md)
>
> Break one forecast into a starting value and the contribution from each
> input.

> [card] [What is the forecast sensitive to?](https://nixtla.io/docs/forecasting/explanation/intervention.md)
>
> See how the forecast changes when one input is replaced by its typical
> historical value.

> [card] [Which signals have been useful historically?](https://nixtla.io/docs/forecasting/exogenous-variables/causal-explanations.md)
>
> Rank the features that have been most useful for predicting your target
> over time.

> [card] [I need more control](https://nixtla.io/docs/forecasting/explanation/advanced-explanations.md)
>
> Compare relationship analyses, history-weighted forecast allocations, and
> stability across data windows.

> [!TIP]
>
> If you have a forecast in front of you and are unsure where to start, choose
> [Explain a forecast with SHAP](https://nixtla.io/docs/forecasting/exogenous-variables/interpretability_with_shap.md).

## One example, three questions

The guides use the same retail example throughout: a store forecasts daily
demand using price, promotion, and temperature.

![A retail-demand forecast beside a bar chart showing price, promotion, and temperature contributions](https://mintcdn.com/nixtla-enterprise/Z27JIqRDnQ0rRCKY/images/forecasting/explain-retail-shap.png?fit=max\&auto=format\&n=Z27JIqRDnQ0rRCKY\&q=85\&s=0bfb97054271fa0cbe8bd5eb1777b6c9)

An actual TimeGPT 2.1 forecast and its feature contributions for one promotion day.

Each guide looks at this example from a different angle:

| Question | What you learn |
| - | - |
| Why did TimeGPT forecast 92 units on this day? | How price, promotion, and temperature moved that forecast up or down |
| What changes when an input is set to a typical value? | Which inputs the forecast is most sensitive to over the next 14 days |
| Which signals have been useful over the last year? | The relative historical strength of price, promotion, and temperature |

## A simple rule

* Use **SHAP** to explain a forecast you already made.
* Use **intervention** to explore the forecast's sensitivity.
* Use **historical signals** before forecasting or when reviewing your data.

> [!NOTE]
>
> Explanations are a practical way to investigate forecasts and prioritize
> follow-up work. For a major business change, combine them with your normal
> testing and decision process.

## Next

Start with [Explain a forecast with
SHAP](https://nixtla.io/docs/forecasting/exogenous-variables/interpretability_with_shap.md), or go
directly to [Test forecast
sensitivity](https://nixtla.io/docs/forecasting/explanation/intervention.md).