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

# Causal Decomposition

> How Rekko breaks down probability into explainable causal factors.

Causal decomposition is what makes Rekko different from a simple odds aggregator. Instead of just giving you a number, Rekko explains **why** it estimates a given probability by breaking it into weighted causal factors with Bayesian updates.

## What you get

Request `?expand=causal` on any analysis or signal endpoint:

```json theme={null}
{
  "causal": {
    "overall_probability": 0.74,
    "overall_confidence": 0.73,
    "method": "weighted_bayesian",
    "factors": [
      {
        "claim": "Recent inflation data supports a rate cut",
        "direction": "supports_yes",
        "weight": 0.35,
        "confidence": 0.82,
        "prior": 0.50,
        "posterior": 0.78,
        "evidence": [
          "CPI fell to 2.1% in February, below 2.3% forecast",
          "Core PCE at 2.4%, trending toward 2% target"
        ]
      },
      {
        "claim": "Labor market remains tight",
        "direction": "supports_no",
        "weight": 0.25,
        "confidence": 0.71,
        "prior": 0.50,
        "posterior": 0.35,
        "evidence": [
          "Unemployment at 3.7%, near historic lows",
          "Wage growth at 4.1% YoY"
        ]
      },
      {
        "claim": "Fed forward guidance signals openness to cuts",
        "direction": "supports_yes",
        "weight": 0.40,
        "confidence": 0.68,
        "prior": 0.50,
        "posterior": 0.82,
        "evidence": [
          "Powell's February testimony noted 'progress on inflation'",
          "Dot plot median shifted down by 25bps"
        ]
      }
    ],
    "generated_at": "2026-03-24T14:30:00Z"
  }
}
```

## Factor anatomy

Each `CausalFactor` represents one driver of the outcome:

| Field        | Type        | Description                                               |
| ------------ | ----------- | --------------------------------------------------------- |
| `claim`      | string      | What this factor asserts                                  |
| `direction`  | string      | `"supports_yes"`, `"supports_no"`, or `"neutral"`         |
| `weight`     | float (0-1) | Relative importance — weights across factors sum to \~1.0 |
| `confidence` | float (0-1) | How confident Rekko is in this specific factor            |
| `prior`      | float (0-1) | Base rate probability before evidence                     |
| `posterior`  | float (0-1) | Updated probability after incorporating evidence          |
| `evidence`   | string\[]   | Key evidence supporting this factor                       |

## How it works

1. **Factor identification** — The research pipeline identifies 3-7 independent causal drivers of the outcome
2. **Prior estimation** — Each factor starts with a base rate (prior) drawn from historical patterns
3. **Evidence gathering** — Deep web research collects relevant evidence for each factor
4. **Bayesian update** — Each factor's prior is updated to a posterior based on the evidence strength
5. **Weighted aggregation** — Factors are combined using their weights to produce the overall probability

The `method` field indicates the aggregation approach:

| Method              | Description                                      |
| ------------------- | ------------------------------------------------ |
| `weighted_bayesian` | Standard weighted Bayesian combination (default) |
| `linear`            | Simple weighted average of posteriors            |
| `log_odds`          | Log-odds weighted combination                    |

## Why this matters

Most prediction market APIs just give you a price. Rekko gives you the **reasoning chain**:

* **Audit the logic** — See exactly which factors drive the estimate and whether you agree
* **Identify blind spots** — Check if important factors are missing from the decomposition
* **Track changes** — When you re-analyze later, see which factors shifted and why
* **Override confidently** — If you have domain expertise on a specific factor, you can judge whether the overall estimate is too high or too low

## Requesting causal decomposition

Add `?expand=causal` to any analysis or signal endpoint:

```bash theme={null}
# On an analysis
curl "https://api.rekko.ai/v1/markets/kalshi/KXFED-26MAR19/analysis?expand=causal" \
  -H "Authorization: Bearer YOUR_API_KEY"

# On a signal
curl -X POST https://api.rekko.ai/v1/signals \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"platform": "kalshi", "market_id": "KXFED-26MAR19"}' \
  | jq '.causal.factors[] | {claim, direction, weight, posterior}'
```