T0_modeling_experimentation-canary-errors-v1

Per-target, per-horizon prediction errors for the T0 selection placebo (canary, 10 seeds x 2 sigma_z x 5 arms x 2 targets x 14 methods x 26 horizons).

Dataset Info

• Rows: 72800
• Columns: 23

Columns

Column	Type	Description
sigma_z	Value('float64')	loading score std σ_z of the panel's DGP (1.0 or 2.0)
seed	Value('int64')	panel seed; the same seed is the same panel (y, m, q, u) in every arm
arm	Value('large_string')	selection arm: A0 fully random, A random unit with latent-timed T0, B selection on own loadings, C selection on own AR(1) shock, F fixed reference (units 0..k-1 at a fixed T0)
method	Value('large_string')	predictor key
method_class	Value('large_string')	convex
forecast_basis	Value('large_string')	donors (target history only fits weights)
unit	Value('int64')	index of the target unit in the panel (0..15)
t0	Value('int64')	0-based row of T0; fit rows [t0-64, t0), predicted rows [t0, t0+26)
q	Value('float64')	target's per-unit signal share q_i ~ Beta(18, 2) clipped to [0.05, 0.95]
h	Value('int64')	horizon, 1..26; h = 1 is row t0 (the row that triggered selection)
pred	Value('float64')	prediction in raw units
truth	Value('float64')	headline counterfactual m + u (noiseless signal plus the unit's own shock) at row t0+h-1
y	Value('float64')	observed outcome m + u + eps at row t0+h-1
abs_err	Value('float64')
signed_err	Value('float64')	pred - truth (negative = under-prediction)
abs_err_y	Value('float64')
signed_err_y	Value('float64')	pred - y
d	Value('float64')	MASE denominator: mean
target_u	Value('float64')	target's shock u at row t0+h-1 (raw units)
target_z	Value('float64')	target's standardised shock u / sd_pop(u) at row t0+h-1
donor_u_w	Value('float64')	Σ_j w_j u_j over the method's donor weights at row t0+h-1 (the donors' shock contribution to the prediction; for repo simplex the weights sum to 1 - constant); NaN for methods without donor weights (tasc, naive_last, ar_p)
donor_u_mean	Value('float64')	unweighted mean of the donors' u at row t0+h-1
runtime_s	Value('float64')	fit + predict wall time, seconds

Generation Parameters

{
  "script_name": "t0_modeling/minimal_placebo.py",
  "model": "none (CPU simulation; classical SC estimators)",
  "experiment_name": "T0_modeling_experimentation",
  "job_id": "local:canary",
  "cluster": "local (macOS laptop)",
  "artifact_status": "final",
  "canary": true,
  "hyperparameters": {
    "dgp": {
      "n_donors": 15,
      "T_pre": 96,
      "T_post": 32,
      "n_latents": 4,
      "latent_mode": "kernelsynth",
      "loading_family": "simplex",
      "mixing_mode": "static",
      "tau_max": 0,
      "signal_share": 0.9,
      "noise_model": "per_unit",
      "share_concentration": 20.0,
      "noise_floor": 0.001,
      "shock_ratio": 0.25,
      "shock_phi": 0.8
    },
    "score": {
      "gamma": 0.9,
      "treated_share": 0.1,
      "window_share": 0.5,
      "deadline_share": 0.8
    },
    "arms": {
      "A0": {
        "rho": 0.0,
        "beta": 0,
        "score_latents": "independent",
        "gamma": 0.9,
        "treated_share": 0.1,
        "window_share": 0.5,
        "deadline_share": 0.8,
        "centre": "pre_window"
      },
      "A": {
        "rho": 0.0,
        "beta": 0,
        "score_latents": "panel",
        "gamma": 0.9,
        "treated_share": 0.1,
        "window_share": 0.5,
        "deadline_share": 0.8,
        "centre": "pre_window"
      },
      "B": {
        "rho": 1.0,
        "beta": 0,
        "score_latents": "panel",
        "gamma": 0.9,
        "treated_share": 0.1,
        "window_share": 0.5,
        "deadline_share": 0.8,
        "centre": "pre_window"
      },
      "C": {
        "rho": 0.0,
        "beta": 1,
        "score_latents": "panel",
        "gamma": 0.9,
        "treated_share": 0.1,
        "window_share": 0.5,
        "deadline_share": 0.8,
        "centre": "pre_window"
      }
    },
    "sigma_z": [
      1.0,
      2.0
    ],
    "seeds": "10 (0..9)"
  },
  "code_commit": "b595920ef04d96e8151712a32fd4da333f4d3bab",
  "input_datasets": [],
  "reproduce": "git checkout b595920ef04d96e8151712a32fd4da333f4d3bab && python -m t0_modeling.minimal_placebo --seeds 10 --out <dir>",
  "description": "Per-target, per-horizon prediction errors for the T0 selection placebo (canary, 10 seeds x 2 sigma_z x 5 arms x 2 targets x 14 methods x 26 horizons)."
}

Usage

from datasets import load_dataset

dataset = load_dataset("meganrichards3/T0_modeling_experimentation-canary-errors-v1", split="train")
print(f"Loaded {len(dataset)} rows")

---