{"id":2434,"date":"2025-12-02T10:39:31","date_gmt":"2025-12-02T10:39:31","guid":{"rendered":"https:\/\/blogs.mathworks.com\/finance\/?p=2434"},"modified":"2025-12-02T13:06:41","modified_gmt":"2025-12-02T13:06:41","slug":"speeding-up-dynare-models-practical-paths-to-performance-gains","status":"publish","type":"post","link":"https:\/\/blogs.mathworks.com\/finance\/2025\/12\/02\/speeding-up-dynare-models-practical-paths-to-performance-gains\/","title":{"rendered":"Speeding Up Dynare Models: Practical Paths to Performance Gains"},"content":{"rendered":"

Dynamic Stochastic General Equilibrium (DSGE) models are essential tools for policy analysis and forecasting, but estimation runs often exceed 24 hours\u2014particularly for large-scale models or Bayesian approaches. This blog presents three practical techniques for accelerating Dynare workflows using the parallel computing capabilities in MATLAB. These methods deliver measurable performance gains with minimal changes to existing code.<\/p>\n

Three Ways to Improve Performance<\/h1>\n
    \n
  1. Parallelise optimization: <\/strong>Use built-in flags to compute gradients concurrently during likelihood-based estimation.<\/li>\n
  2. Accelerate Bayesian estimation: <\/strong>Distribute independent Markov Chain Monte Carlo (MCMC) across workers for Metropolis-Hastings estimation.<\/li>\n
  3. Manage large experiment grids: <\/strong>Use Experiment Manager to organise and audit multiple calibrations without manual loops.<\/li>\n<\/ol>\n

    Each strategy is examined in detail below, with implementation guidance for Dynare workflows.<\/p>\n

    Parallelizing Optimization<\/strong><\/h1>\n

    Estimating a DSGE model typically involves solving a complex optimization problem to identify parameter values that maximize the likelihood function. Dynare supports various algorithms for this task, including the MATLAB fmincon and other built-in routines.<\/p>\n

    Most of these algorithms rely on gradient-based methods, which require computing derivatives of the objective function. This step is computationally expensive but highly parallelizable, making it an ideal candidate for performance improvement.<\/p>\n

    \"\"<\/p>\n

    MATLAB users may be familiar with the UseParallel<\/tt> flag, which calculates directional gradients in parallel to reduce bottlenecks in high-dimensional problems. A natural approach would be to enable this in Dynare as follows:<\/p>\n

    estimation(optim=('UseParallel', true));<\/code><\/p>\n

    However, as of Dynare 6, this approach will produce incorrect results when used without additional configuration. The issue stems from Dynare’s persistent internal state, which does not automatically transfer to parallel workers.<\/p>\n

    The solution requires carrying Dynare’s state into the parallel environment\u2014a process that is not immediately obvious. To address this, we have published a utility function on GitHub that handles the state transfer automatically:<\/p>\n

    dynareParallel(\u201cmyModel.mod\u201d)<\/code><\/p>\n

    This function ensures that parallel execution is both correct and efficient while preserving the standard Dynare syntax.<\/p>\n

    Performance tip:<\/strong> Match the number of workers to the number of optimisation variables. Using 100 workers for 10 variables wastes resources.<\/p>\n

    Accelerating Bayesian Estimation<\/strong><\/h1>\n

    Bayesian estimation with Metropolis-Hastings (MH) represents another significant computational bottleneck in Dynare workflows. MH sampling can be configured with multiple chains and blocks:<\/p>\n

    estimation(mh_replic = 20000, mh_nblocks = 4);<\/code><\/p>\n

    Because each chain is independent, this configuration is well suited for parallel execution. Distributing four chains across four workers can reduce runtime by a factor of four\u2014for example, reducing a 40-hour estimation to approximately 10 hours with minimal code changes.<\/p>\n

    As with gradient-based optimization, parallelizing MH chains requires proper handling of Dynare’s internal state when using the MATLAB Parallel Computing Toolbox. The same dynareParallel<\/tt> helper function introduced above can be used for both Dynare 5 and 6. This function distributes chains across workers and ensures correct state transfer.<\/p>\n

     <\/p>\n

    \"\"<\/p>\n

     <\/p>\n

    Managing Large Experiment Grids<\/strong><\/h1>\n

    In some cases, the challenge is not a single long-running estimation but rather hundreds of smaller simulations for testing different calibrations or policy scenarios.<\/p>\n

    The MATLAB Experiment Manager app provides a structured framework for managing these workflows. The tool enables users to:<\/p>\n