Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →EconML (Microsoft) — heterogeneous treatment effect estimation. Double ML, Causal Forest, Deep IV, and metalearners (S-Learner, T-Learner, X-Learner). Orthogonal learning for causal effects from observational data.
.claude/skills/mkurman-econml/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 51% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 50% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -27% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -18% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -49% | 0% |
EconML is a Microsoft library for causal inference and heterogeneous treatment effect estimation using machine learning. Implements Double ML, Causal Forest, DML, IV methods, and orthogonal statistical learning. Designed for observational data where treatment effects vary across individuals.
bashuv pip install econml
pythonfrom econml.dml import LinearDML import numpy as np X = np.random.randn(500, 5) # features T = np.random.randn(500) # treatment Y = T * (0.5 + X[:, 0]) + np.random.randn(500) # outcome est = LinearDML(model_y="auto", model_t="auto", discrete_treatment=False) est.fit(Y, T, X=X) print(f"ATE: {est.ate():.3f} ± {est.ate_inference().stderr:.3f}")
pythonfrom econml.grf import CausalForest cf = CausalForest(n_estimators=100, min_samples_leaf=10) cf.fit(X, T, Y) treatment_effects = cf.effect(X) print(f"Heterogeneous effects range: {treatment_effects.min():.3f} to {treatment_effects.max():.3f}")
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 12,232 | 9,636 | -21% | 1 | 1 | 0% | 2,376 | 2,396 | +1% | 0 | 0 | — |
case-02 | fail→fail | 12,387 | 7,921 | -36% | 1 | 1 | 0% | 2,462 | 1,780 | -28% | 0 | 0 | — |
case-03 | fail→fail | 12,764 | 9,315 | -27% | 1 | 1 | 0% | 2,428 | 2,294 | -6% | 0 | 0 | — |
case-04 | fail→pass | 3,288 | 2,415 | -27% | 1 | 1 | 0% | 502 | 760 | +51% | 0 | 0 | — |
case-05 | pass→pass | 8,172 | 2,446 | -70% | 1 | 1 | 0% | 1,545 | 798 | -48% | 0 | 0 | — |
case-06 | fail→pass | 5,548 | 5,237 | -6% | 1 | 1 | 0% | 939 | 1,410 | +50% | 0 | 0 | — |
case-07 | pass→pass | 4,180 | 2,572 | -38% | 1 | 1 | 0% | 772 | 836 | +8% | 0 | 0 | — |
case-08 | fail→pass | 13,484 | 7,406 | -45% | 1 | 1 | 0% | 2,302 | 1,683 | -27% | 0 | 0 | — |
case-09 | fail→pass | 13,963 | 8,396 | -40% | 1 | 1 | 0% | 2,612 | 2,141 | -18% | 0 | 0 | — |
case-10 | pass→pass | 11,170 | 9,024 | -19% | 1 | 1 | 0% | 2,152 | 2,078 | -3% | 0 | 0 | — |
case-11 | pass→pass | 7,285 | 3,375 | -54% | 1 | 1 | 0% | 1,282 | 960 | -25% | 0 | 0 | — |
case-12 | pass→pass | 2,517 | 1,560 | -38% | 1 | 1 | 0% | 356 | 563 | +58% | 0 | 0 | — |
case-13 | pass→pass | 7,541 | 3,746 | -50% | 1 | 1 | 0% | 1,456 | 991 | -32% | 0 | 0 | — |
case-14 | pass→pass | 1,737 | 2,396 | +38% | 1 | 1 | 0% | 231 | 557 | +141% | 0 | 0 | — |
case-15 | fail→pass | 10,940 | 4,086 | -63% | 1 | 1 | 0% | 2,073 | 1,054 | -49% | 0 | 0 | — |
case-16 | pass→pass | 9,411 | 4,852 | -48% | 1 | 1 | 0% | 1,624 | 1,202 | -26% | 0 | 0 | — |
case-17 | pass→fail | 6,780 | 3,636 | -46% | 1 | 1 | 0% | 1,265 | 1,042 | -18% | 0 | 0 | — |
case-18 | fail→pass | 5,597 | 3,748 | -33% | 1 | 1 | 0% | 1,000 | 898 | -10% | 0 | 0 | — |
case-19 | pass→pass | 18,813 | 16,191 | -14% | 1 | 1 | 0% | 3,390 | 3,432 | +1% | 0 | 0 | — |
case-20 | pass→pass | 8,671 | 6,217 | -28% | 1 | 1 | 0% | 1,714 | 1,470 | -14% | 0 | 0 | — |
case-21 | pass→pass | 14,611 | 11,201 | -23% | 1 | 1 | 0% | 2,729 | 2,537 | -7% | 0 | 0 | — |
case-22 | pass→pass | 10,409 | 8,131 | -22% | 1 | 1 | 0% | 2,130 | 2,089 | -2% | 0 | 0 | — |
case-23 | pass→pass | 2,735 | 2,282 | -17% | 1 | 1 | 0% | 456 | 696 | +53% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +22 percentage points is the difference between those two pass rates over the 23 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.