Explain metric shift shapley

What it's for

Makes Gitrevio's 'why did this metric move?' and Shapley claims mathematically inspectable while preventing a descriptive decomposition from being marketed as causal attribution.

What you give it

Every tool is a pure function: it computes from what you pass it and reads nothing on its own. Inputs split into evidence you supply from your connected systems, calibration your team owns, and numerical controls that affect precision but never the result's meaning.

{
  "bootstrap_draws": 200,
  "current_observations": [
    {
      "features": {
        "review_load": 2,
        "rework_rate": 0
      },
      "id": "shapley-current-0",
      "outcome": 4
    },
    {
      "features": {
        "review_load": 3,
        "rework_rate": 3
      },
      "id": "shapley-current-1",
      "outcome": 7.5
    },
    {
      "features": {
        "review_load": 4,
        "rework_rate": 6
      },
      "id": "shapley-current-2",
      "outcome": 11
    },
    {
      "features": {
        "review_load": 5,
        "rework_rate": 2
      },
      "id": "shapley-current-3",
      "outcome": 11
    },
    {
      "features": {
        "review_load": 6,
        "rework_rate": 5
      },
      "id": "shapley-current-4",
      "outcome": 14.5
    },
    {

What you get back

This is the actual output of running the example above — computed by the same function the platform calls, not an illustration.

{
  "assumptions": [
    "The pooled second-order response surface predicts the declared aggregate metric adequately out of sample.",
    "Reference/current cohorts, feature definitions, and missingness policies are comparable.",
    "The decomposition allocates modeled association, including half of each pair interaction to each participant; it is not a causal Shapley effect."
  ],
  "attribution_count": 2,
  "attributions": [
    {
      "confidence_interval": [
        2.1472,
        6.1517
      ],
      "contribution": 3.9995,
      "feature": "review_load",
      "same_direction_probability": 1,
      "share_of_observed_shift": 0.9999
    },
    {
      "confidence_interval": [
        -0.3506,
        0.4125
      ],
      "contribution": 0,
      "feature": "rework_rate",
      "same_direction_probability": 0.51,
      "share_of_observed_shift": 0
    }
  ],
  "configuration": {
    "bootstrap_draws": 200,
    "confidence_level": 0.9,
    "maximum_unexplained_fraction": 0.5,
    "minimum_absolute_shift": 0.1,
    "ridge_penalty": 0.01
  },
  "decision": "metric_shift_decomposed",
  "diagnostics": {
    "bootstrap_fallback_draws": 0,
    "decomposition_efficiency_error": 0,
    "minimum_out_of_sample_r_squared": 0.05,
    "model_term_count": 3,
    "observed_shift_confidence_interval": [
      2.2275,

How it works

Statistical audit & measurement — Check whether a number is fit to decide on: coverage, timing, reconciliation, and the gaps a dashboard hides.

1. 1 Pool comparable reference and current observations, standardize declared factors, and fit a ridge response surface containing main effects and optional pair interactions.
2. 2 Require stratified five-fold out-of-sample predictive fit before treating the response surface as explanatory.
3. 3 Decompose the fitted mean shift into main-effect contributions and split every interaction contribution equally between its participating factors so contributions exactly sum to the modeled shift.
4. 4 Bootstrap observations independently within each window, retain uncertainty and sign stability for every contribution, and keep the actual-minus-modeled shift visible as an unexplained residual gate.

Before you trust it

Every tool in the catalog ships with the conditions under which its answer is meaningful — and the conditions under which it should abstain instead of guessing.

Calibrating it to your org

Calibration workflow

1. 1 Define the management decision, target outcome, aggregate unit, privacy boundary, cadence, and prediction/intervention horizon for this organization.
2. 2 Build a tenant-scoped historical cohort using only information available before each prediction or decision; preserve zero periods, censoring, assignment probabilities, and unresolved outcomes when the method requires them.
3. 3 Estimate statistical parameters on training history, but obtain costs, utilities, risk tolerance, practical-effect thresholds, capacity, and policy constraints from accountable decision owners.
4. 4 Validate on later time windows or held-out aggregate units at the deployment grain, against a simple baseline and the function-specific validation strategy.
5. 5 Deploy only if the returned decision clears evidence, overlap, calibration, robustness, and guardrail checks; warning, unsupported, schema-gap, and fallback decisions are abstentions.
6. 6 Monitor realized outcomes, data drift, coverage, and decision regret; recalibrate at a governed cadence or after a detected regime/definition change, never merely because a stakeholder dislikes the result.

Call it from your AI

You don't wire up 388 tools in your MCP client. The GitRevio MCP server exposes 20 tools, three of which let an agent search the catalog, read a tool's schema, and run it — so the assistant finds this one on its own.

gitrevio_capabilities_search
  { "q": "fit a crossvalidated secondorder ridge response" }
  → finds "explain_metric_shift_shapley"

gitrevio_capability_describe
  { "capability_id": "explain_metric_shift_shapley" }
  → returns the input schema and agent guidance shown on this page

gitrevio_capability_run
  { "capability_id": "explain_metric_shift_shapley", "arguments": { ... } }
  → returns the result shown above

Works in Claude Desktop, Claude Code, Cursor, Cline, Continue.dev, Goose and Aider. See the MCP server.

Related tools

See every tool in People, retention & knowledge →