8. Computational Speed and Resources for Effect Estimation

Training and prediction have different resource demands. Adjust settings with the estimand and the retained population in mind; reducing computation can also change the statistical result.

8.1. Parallel Processing

For parallel processing with Ray, use Python 3.12 on Windows, or Python 3.13 on Linux and macOS. Joblib is installed as a dependency of mcf and also supports Python 3.14. See Installation Guide.

By default, ModifiedCausalForest uses Joblib when Ray is unavailable or when the adjusted training sample is sufficiently small on Windows. Otherwise, it uses Ray. See _int_mp_backend in the API for the exact selection rule.

gen_mp_parallel controls the number of parallel processes. Reducing this number can lower memory use. If parallel processing causes problems, create a new instance with _int_mp_backend='joblib' to use Joblib, or _int_mp_backend='sequential' to run sequentially. You can also set gen_mp_parallel=1 to use a single process. When explicitly selecting Joblib on Windows, the initial number of workers must not exceed 63.

8.2. Forest Tuning via Grid Search

Grid search considers three parameters: candidate-variable counts, alpha balance and minimum leaf sizes. The forest evaluates combinations using its out-of-bag objective.

Parameter	Effect on computation
cf_boot	Number of trees; more trees require more computation.
cf_m_share_min, cf_m_share_max	Bounds for the share of candidate variables.
cf_m_grid	Number of candidate-count grid values, not variables per split.
cf_alpha_reg_min, cf_alpha_reg_max	Split-balance bounds; discrete treatments use treatment-specific balance by default.
cf_n_min_min, cf_n_min_max	Minimum-leaf-size search range.
cf_tune_all	Expands tuning grids; increases computation.
cf_random_thresholds	Limits initial split candidates. If no admissible split is found, the algorithm retries all available variables and supported thresholds.
cf_chunks_maxsize	Splits large training samples into chunks and averages the resulting forests.
cf_subsample_power_eval	Filling-sample exponent, default 0.90; construction uses 0.85.
cf_subsample_factor_eval	None/True resolves to factor 2; False uses all available filling observations.

See Training a Modified Causal Forest for filling and candidate-variable rules. Binary features should be listed as ordered variables.

8.3. Minimization of RAM usage

Low-memory prediction is enabled by default. It processes forest weights in chunks rather than retaining the full weight matrix. Consider chunking and fewer workers before discarding prediction observations. QIATEs and p_iate_m_ate require _int_low_memory_predict=False.

The prediction-row cap is 1,000,000 with low-memory prediction and 250,000 otherwise, unless changed through _int_max_obs_prediction. A smaller prediction sample can change the population over which effects are averaged; it is not solely a performance setting. Training-sample reductions can also reduce precision.

For BGATE and CBGATE estimation, mcf draws a reference sample from the prediction data and reuses it at each value of the heterogeneity variable being evaluated. At least 50 observations are used per evaluation value, or all available observations if fewer than 50 remain. As a result, the total number of reference rows can exceed the number implied by p_bgate_sample_share. See CBGATEs and BGATEs for details. p_choice_based_sampling reweights aggregate prediction observations by treatment. It does not draw a subsample to speed up computation.

8.4. Example

This configuration fragment illustrates resource controls. See the API for sample-dependent defaults and the remaining options.

from mcf.mcf_main import ModifiedCausalForest

my_mcf = ModifiedCausalForest(
    var_y_name='outcome', var_d_name='treat',
    var_x_name_ord=['x_cont0', 'x_cont1', 'x_ord1'],
    cf_boot=500,
    gen_mp_parallel=2,
    _int_low_memory_predict=True
)

GPU-related requirements are described in Experimental features.