Statistical Factor Models Server

Overview

This server fits statistical factor models to a supplied numeric panel. It makes no external data-provider calls.

Model	Tool prefix	Panel
Complete-panel PCA with Bai-Ng rank selection	`complete-panel-factor-extractor`	Balanced, no missing cells
Stock-Watson EM	`swfactor-extractor`	Missing observations and explicit mixed-frequency mappings
Banbura-Modugno dynamic factor model	`bmfactor-extractor`	Missing observations, dynamic/state-space modeling

Connection

MCP URL: https://quantx-api.limex.pro/mcps/statistical-factor-models Select Statistical Factor Models in the QuantX marketplace. Use its gateway connection and QuantX authorization; no market-data or broker key is required. See Client Setup. Cursor configuration (.cursor/mcp.json):

{
  "mcpServers": {
    "statistical-factor-models": {
      "url": "https://quantx-api.limex.pro/mcps/statistical-factor-models",
      "headers": {
        "Authorization": "Bearer ${env:QUANTX_MCP_TOKEN}"
      }
    }
  }
}

Set the referenced environment variables before starting Cursor. For Claude Code or Python, use the client-specific format in Client Setup.

Tools

Tools	Inputs and purpose	Result
`complete-panel-factor-extractor-fit`	Required `X`, optional configured `handle`; fit a balanced panel	Factors, loadings, reconstructed panel, rank-selection metadata
`swfactor-extractor-fit`	Required `handle` and `X`; fit a panel with gaps	Factors, loadings, reconstruction and EM diagnostics
`bmfactor-extractor-fit`	Required `X`, optional configured `handle`; dynamic factor fit	Factors, loadings, reconstruction and state-space diagnostics

Each prefix also has -create, -info, -list, and -delete tools. Use types-list to inspect constructors; place model settings in init_kwargs of -create, not in -fit. For reproducible settings, create an explicit handle. Pass exactly one of n_factors or n_factors_max. Rank settings and SW/BM max_iter must be positive JSON integers, not strings, booleans, or floats. Automatic rank selection requires n_factors_max < min(N, T).

Examples

Fit a Balanced Panel

Call complete-panel-factor-extractor-create:

{
  "request": {
    "init_kwargs": {
      "n_factors": 1
    }
  }
}

Then call complete-panel-factor-extractor-fit with the returned handle. This is synthetic stationary input, with rows as series and columns as observations:

{
  "request": {
    "handle": "<handle-from-create>",
    "X": [
      [0.01, -0.02, 0.015, -0.005, 0.012, -0.008],
      [0.02, -0.01, 0.010, -0.003, 0.018, -0.012],
      [0.005, -0.015, 0.012, -0.002, 0.009, -0.004]
    ]
  }
}

Fit a Panel With Gaps

Call swfactor-extractor-create:

{
  "request": {
    "init_kwargs": {
      "n_factors": 1,
      "max_iter": 100,
      "flow_series": {},
      "flow_series_diff": {}
    }
  }
}

Then call swfactor-extractor-fit:

{
  "request": {
    "handle": "<handle-from-create>",
    "X": [
      [0.01, null, 0.015, -0.005, 0.012, -0.008],
      [0.02, -0.01, null, -0.003, 0.018, -0.012],
      [0.005, -0.015, 0.012, -0.002, 0.009, -0.004]
    ]
  }
}

Use the bmfactor-extractor-create / bmfactor-extractor-fit pair for the dynamic model and inspect its constructor rather than reusing SW-specific arguments.

Results and Limitations

Input shape is (N, T): series by observations. Do not pass a date-by-ticker matrix without transposing it.
Rows should already be stationary, for example returns or appropriately differenced series. A model does not make arbitrary price levels stationary for you.
Fit payloads include factors_ with shape (k, T), loadings_ with shape (N, k), and X_fitted_ with shape (N, T), plus model-specific diagnostics. Resolve result_id when the payload is stored.
JSON null is supported for SW/BM missing observations. Complete-panel PCA rejects it. SW and BM also need at least one observed value in every row and every column; Banbura-Modugno additionally needs at least three time points (T >= 3). Do not replace gaps with zero unless that is a deliberate modeling choice.
Classify SW mixed-frequency rows by the stationary series actually supplied: ordinary missing/stock rows do not belong in flow mappings; flow_series and flow_series_diff describe different aggregation assumptions. Inspect the constructor guidance before setting them.
Retain the returned factors and diagnostics separately. A later call can reuse the fitted object while it remains loaded.

See Shared MCP Tools for importing tables without losing missing cells and for exporting fit results.

Overview

MCP Setup

Dataloader Server

Fama-French Replicate Server