Statistical Factor Models Server

Overview

The Statistical Factor Models Server fits statistical factor models from numeric panel matrices. It covers Stock-Watson EM, complete-panel Bai-Ng style extraction, and Banbura-Modugno dynamic factor extraction.

Connection

Add this server to your MCP client configuration.

Cursor
Bearer auth

{
  "mcpServers": {
    "statistical-factor-models": {
      "url": "https://backend-production-c775c.up.railway.app/mcp/statistical-factor-models"
    }
  }
}

{
  "mcpServers": {
    "statistical-factor-models": {
      "url": "https://backend-production-c775c.up.railway.app/mcp/statistical-factor-models",
      "headers": {
        "Authorization": "Bearer ${env:STATISTICAL_FACTOR_MODELS_MCP_TOKEN}"
      }
    }
  }
}

Restart the client after changing MCP configuration. The server tools appear automatically after the connection is established.

Transport

Property	Value
Protocol	MCP over Streamable HTTP
MCP URL	`https://backend-production-c775c.up.railway.app/mcp/statistical-factor-models`
Health URL	`https://backend-production-c775c.up.railway.app/health`
MCP path	`/mcp/statistical-factor-models`
Request envelope	`{"request": {...}}`
Auth	Optional bearer token when enabled for the endpoint

Best-Fit Workflows

Fit a balanced panel with no missing values.
Fit a missing-value panel with Stock-Watson EM.
Fit a missing-value panel with a dynamic/state-space factor model.
Use explicit mixed-frequency mappings for Stock-Watson runs.

Recommended Tools

complete-panel-factor-extractor-fit
swfactor-extractor-create -> swfactor-extractor-fit
bmfactor-extractor-fit

Tools

Tool	Description	Returns
`swfactor-extractor-fit`	Fit the Stock-Watson EM factor model on a 2D panel `X` with shape `(N, T)`, where rows are series and columns are time points. Missing observations may be passed as `null`/NaN. Every row of `X` must already be stationary / `I(0)`; if the underlying series are `I(1)`, transform them first, for example with first differences or log differences. Preferred agent path: first create a handle with `swfactor-extractor-create`, pass exactly one of `n_factors` or `n_factors_max` in `init_kwargs`, and always pass explicit `flow_series` and `flow_series_diff` there, even when both are empty dicts. Then call `swfactor-extractor-fit` with that `handle`. Do not pass `n_factors_max` directly to `swfactor-extractor-fit`, and do not rely on direct/singleton SW calls to fill `flow_series` or `flow_series_diff` for you. For mixed-frequency setup, ordinary MAR rows and end-of-period stock-style rows should not go into `flow_series` or `flow_series_diff`; use `flow_series[i]=M` for Type C / Type D rows that decompose into `M` higher-frequency `I(0)` pieces, and use `flow_series_diff[i]=M` only for Type E rows built as stationary transforms of averaged `I(1)` levels. Classify rows by the stationary series actually passed into `X`, not by the raw source column. Example: if `X` rows are adjusted log-returns for `[SPY_adj_close, QQQ_adj_close, IWM_adj_close, ...]` and only `SPY` and `QQQ` are quarter-end rows on a monthly grid, use `flow_series={0:3,1:3}` and `flow_series_diff={}`. When the subtask is only to prepare SW mixed-frequency mappings, return only those two arguments.	`Any`
`complete-panel-factor-extractor-fit`	Fit the complete-panel Bai-Ng style factor model on a balanced 2D numeric panel `X` with shape `(N, T)`. Missing values are not allowed. Every row of `X` must already be stationary / `I(0)`; if the underlying series are `I(1)`, transform them first, for example with first differences or log differences. Fast happy path: call this tool directly with `{"request":{"X":[[1.0,2.0,3.0],[1.5,2.5,3.5]]}}`. If you need custom rank settings, first create a handle with `complete-panel-factor-extractor-create` and pass exactly one of `n_factors` or `n_factors_max` in `init_kwargs`.	`Any`
`bmfactor-extractor-fit`	Fit the Banbura-Modugno dynamic factor model on a 2D panel `X` with shape `(N, T)`, where rows are series and columns are time points. Missing observations may be passed as `null`/NaN. Every row of `X` must already be stationary / `I(0)`; if the underlying series are `I(1)`, transform them first, for example with first differences or log differences. Use this when you want a dynamic/state-space model rather than the simpler SW extractor. Fast happy path: call this tool directly with `{"request":{"X":[[1.0,null,3.0],[2.0,2.5,3.5]]}}`. If you need custom rank settings, first create a handle with `bmfactor-extractor-create` and pass exactly one of `n_factors` or `n_factors_max` in `init_kwargs`.	`Any`

Examples

Fit a balanced panel directly

{
  "request": {
    "X": [
      [
        1.0,
        2.0,
        3.0,
        4.0
      ],
      [
        1.5,
        2.5,
        3.5,
        4.5
      ],
      [
        2.0,
        3.0,
        4.0,
        5.0
      ]
    ]
  }
}

Create a Stock-Watson extractor

{
  "request": {
    "init_kwargs": {
      "n_factors_max": 2,
      "flow_series": {},
      "flow_series_diff": {}
    }
  }
}

Fit the created Stock-Watson extractor

{
  "request": {
    "handle": "<handle-from-create>",
    "X": [
      [
        1.0,
        null,
        3.0,
        4.0
      ],
      [
        2.0,
        2.5,
        null,
        4.5
      ],
      [
        0.5,
        1.0,
        1.5,
        2.0
      ]
    ]
  }
}

Notes

Primary fitting tools expect X as a JSON array of arrays with shape (N, T): rows are series, columns are time points.
Input rows should already be stationary / I(0).
Use null for missing observations in the Stock-Watson and dynamic factor tools.
For Stock-Watson mixed-frequency work, create a handle with explicit flow_series and flow_series_diff first.

Client setup

Configure this endpoint in Cursor, Claude Desktop, or a generic MCP client.

Shared tools

Use health, result, artifact, environment, and table helper tools.

Other Servers

Dataloader

Market-data retrieval, TA-Lib indicators, and dataframe exports.

Fama-French Replicate

Official and replicated Fama-French factors plus loadings and alpha estimation.

Jump Models

JumpModel and SparseJumpModel regime fitting, online prediction, and backtesting.

Wavelet Mean Reversion

Wavelet-based mean reversion analysis for financial time series.

Parallax ExtremeHurst

ExtremeHurst signal generation from OHLCV data.

EP Ratio Screener

Fundamental stock screening based on earnings yield and balance-sheet quality.

Volatility Scaling Lab

Volatility targeting, EWMA volatility, Monte Carlo bands, and risk diagnostics.

​Overview

​Connection

​Transport

​Best-Fit Workflows

​Recommended Tools

​Tools

​Examples

​Fit a balanced panel directly

​Create a Stock-Watson extractor

​Fit the created Stock-Watson extractor

​Notes