Jump Models Server

Overview

Jump Models fits discrete or continuous regime models and sparse variants with feature selection. Supply your own feature table, for example from Dataloader prices transformed into returns and volatility features. These tools are not a market-data feed; every example below runs on features you pass in.

Connection

MCP URL: https://quantx-api.limex.pro/mcps/jump-models Select Jump Models in the QuantX marketplace. Use its gateway connection and QuantX authorization; no external data-provider key is needed for supplied features. See Client Setup. Cursor configuration (.cursor/mcp.json):

{
  "mcpServers": {
    "jump-models": {
      "url": "https://quantx-api.limex.pro/mcps/jump-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	Purpose	Result
`jump-model-fit`	Fit `X`; optional `ret_ser`, `feat_weights`, and `sort_by`	Fit response; use prediction tools for states and probabilities
`sparse-jump-model-fit`	Fit `X`; optional `ret_ser` and `sort_by`. Set sparsity in the constructor (`max_feats`)	Sparse fit response; use sparse prediction tools
`jump-model-predict-online`, `sparse-jump-model-predict-online`	Online state assignment using the fitted parameters	State sequence
`jump-model-predict-proba-online`, `sparse-jump-model-predict-proba-online`	Online regime probabilities	Observation-by-state probabilities
`jump-model-predict`, `sparse-jump-model-predict`	Batch state assignment	State sequence
`jump-model-predict-proba`, `sparse-jump-model-predict-proba`	Batch probabilities	Observation-by-state probabilities
`jump-model-backtest`	Returns/prices plus states, probabilities, or a fitted handle and `X`	Strategy rows (state, exposure, returns, equity, turnover) plus summary metrics including `max_drawdown`. No per-row drawdown series
`data-loader-load`	Load an example pickle into a `DataLoader` instance, when the deployment includes it	Populates `DataLoader.X` and `DataLoader.ret_ser`; not a table export
`matplotlib-setting`	Configure plot styling	Styling result

The jump-model, sparse-jump-model, and data-loader prefixes each have -create, -info, -list, and -delete tools. Inspect constructors with types-list. There is one jump-model-backtest tool for both model families, not a separate sparse backtest tool. The Nasdaq-100 example file NDX.pkl is missing from the current hosted deployment, so data-loader-load fails there with a file-not-found error. Even where the file exists, the tool fills an internal DataLoader object instead of returning feature rows, so it is not a way to obtain a table. Pass your own feature table as X.

Examples

Create a Model

Call jump-model-create:

{
  "request": {
    "init_kwargs": {
      "n_components": 2,
      "jump_penalty": 10.0,
      "cont": false,
      "random_state": 42
    }
  }
}

Fit Numeric Features

Call jump-model-fit. This small synthetic panel demonstrates the request shape:

{
  "request": {
    "handle": "<handle-from-create>",
    "X": {
      "index": ["2026-01-05", "2026-01-06", "2026-01-07", "2026-01-08", "2026-01-09", "2026-01-12"],
      "return_z": [-1.1, -0.8, -1.0, 0.9, 1.2, 0.8],
      "volatility_z": [1.0, 0.9, 1.1, -0.8, -0.9, -1.0]
    }
  }
}

X is observations by features. Dates belong in the reserved index, not in a numeric feature column. A rows/columns or matrix/columns object is also supported. A matrix produced on this server by table_to_ndarray can be passed as X: {"result_id": "<matrix-result-id>"} — that id must come from Jump Models. jump-model-fit takes three optional arguments. The sparse fit accepts only ret_ser and sort_by, because it selects features itself:

Argument	Effect
`ret_ser`	A return series aligned with `X`, used to order and interpret the regimes
`feat_weights`	Per-feature weights for the fit
`sort_by`	How the fitted states are ordered — default `"cumret"`, cumulative return of each regime. Pass `null` to leave the states in their raw order, which makes the labels arbitrary between runs

Predict Online States

Call jump-model-predict-online with the same handle and features:

{
  "request": {
    "handle": "<handle-from-create>",
    "X": {
      "index": ["2026-01-13", "2026-01-14"],
      "return_z": [0.7, -0.6],
      "volatility_z": [-0.8, 0.5]
    }
  }
}

Use jump-model-predict-proba-online with the same shape for probabilities. An online assignment is not a guaranteed next-day forecast. Fit preprocessing and model parameters on the training window only.

Backtest Explicit States

Call jump-model-backtest with supplied synthetic returns and states:

{
  "request": {
    "returns": [0.01, -0.02, 0.015, 0.005, -0.004, 0.006],
    "states": [0, 0, 1, 1, 0, 0],
    "state_exposure": {"0": 1.0, "1": 0.0},
    "transaction_cost_bps": 5.0,
    "initial_equity": 1.0
  }
}

Inspect the returned rows to confirm exposure timing and costs. State labels are arbitrary: do not assume state 0 always means a bull regime. Determine the mapping from the fitted training sample, without using future returns. The row’s state sets that same row’s exposure — lag the states yourself if a decision at t should apply only at t+1. metrics.max_drawdown is negative (-0.025 is a 2.5% decline).

Results and Limitations

Prediction arrays and backtests may be stored references; resolve them with result_get. Keep the input window, feature transforms, constructor settings, state mapping, and result IDs with a report. A handle identifies the model you created. Fitted parameters remain available for later predict calls while that object is still loaded. Keep fit and prediction on the same handle, and fit again if a call reports the model is missing or not fitted. Save returned states and backtests separately. See Object Handles. The tools do not provide a complete automatic walk-forward research protocol. Batch predictions and in-sample fits should not be presented as out-of-sample evidence. See Shared MCP Tools for table preparation and artifact export.

Overview

MCP Setup

Dataloader Server

Fama-French Replicate Server

Statistical Factor Models Server

Wavelet Mean Reversion Server

EP Ratio Screener Server

Volatility Scaling Lab Server

Jump Models Server

Overview

Connection

Tools

Examples

Create a Model

Fit Numeric Features

Predict Online States

Backtest Explicit States

Results and Limitations

​Overview

​Connection

​Tools

​Examples

​Create a Model

​Fit Numeric Features

​Predict Online States

​Backtest Explicit States

​Results and Limitations

Overview

Connection

Tools

Examples

Create a Model

Fit Numeric Features

Predict Online States

Backtest Explicit States

Results and Limitations