Run in Colab Enterprise

Run in Colab Enterprise: click the badge to import this notebook, pick a runtime, and Run all. The Terraform-deployed templates already carry the SF_* run identity in their env, so there's no environment cell to fill in. Runs on the sf-main runtime template (Python 3.11). See docs/notebook_runtimes.md for the per-notebook template mapping and the headless acceptance harness.

01 · Serverless BigQuery ML Forecasting (ARIMA_PLUS, ARIMA_PLUS_XREG & TimesFM)¶

Execute enterprise time-series forecasting directly inside Google Cloud's serverless data warehouse using pure BigQuery SQL — comparing classical univariate and exogenous state-space decomposition (ML.FORECAST with ARIMA_PLUS and ARIMA_PLUS_XREG + ML.EXPLAIN_FORECAST Two-Tier attributions) against zero-shot foundation model inference (AI.FORECAST with Google Research's TimesFM 2.5 univariate and TimesFM 3.0 multivariate covariates), blending them into a consensus ensemble, and running anomaly, structural change-point, and inventory replenishment diagnostics.

At a glance¶

Dimension	Details
Learning outcomes	Run in-warehouse SQL forecasting (ARIMA_PLUS & ARIMA_PLUS_XREG via ML.FORECAST / ML.EXPLAIN_FORECAST, plus TimesFM 2.5 & TimesFM 3.0 via AI.FORECAST), evaluate 21 metrics + Covariate FVA across rolling-origin folds, inspect Two-Tier ML.EXPLAIN_FORECAST attributions, detect anomalies & structural breaks (detect_anomalies_df, detect_change_points_df), and compute $(s, S)$ inventory replenishment policies (inventory_policy_df).
Prerequisites	Deployed scale-forecasting Google Cloud project (SF_* env vars) + scale-forecasting[gcp].
Estimated time	~3–5 minutes.
Estimated run cost	Estimated ~\$0.05–\$0.20 USD (BigQuery query bytes processed / slot-seconds + BigQuery ML model creation & AI.FORECAST inference).
Ongoing cost	Zero compute when idle (only standard BigQuery storage for the fitted ARIMA_PLUS / ARIMA_PLUS_XREG models and registry rows; delete anytime with python -m scale_forecasting.registry.ops drop-run <RUN_ID> --yes).

[!NOTE] Why Both arima_plus (ARIMA_PLUS / ARIMA_PLUS_XREG) and timesfm (TimesFM 2.5 / 3.0) Are Per-Series (local) Models: Even though BigQuery executes all series in a single vectorized SQL query (TIME_SERIES_ID_COL='ts_id' or id_cols=['ts_id']), each series is forecast in mathematical isolation with zero cross-series parameter sharing (supports_global = False). All 21 evaluation metrics are scored by the exact same Python metric engine (metrics.compute_metrics) used by Spark, Ray, Vertex, GCE, and GKE.

flowchart LR
    subgraph Source["BigQuery / BigLake Iceberg"]
        S1["source_series_covariates_native
or source_series_covariates_iceberg"]
    end
    subgraph NativeEngine["engines/bigquery_engine.py (Serverless SQL)"]
        A1["arima_plus & arima_plus@xreg
CREATE MODEL ... ARIMA_PLUS / ARIMA_PLUS_XREG
+ ML.FORECAST + ML.EXPLAIN_FORECAST"]
        T1["timesfm@univariate & timesfm
AI.FORECAST(TimesFM 2.5 & TimesFM 3.0)"]
        SCORE["_score_fold()
Identical 21-Metric Panel"]
    end
    subgraph Ens["ensemble_run.py + Diagnostics"]
        E1["combine_calculated() · fit_learned()
mean · inverse_error · nnls · champion
+ FVA · Anomalies · Inventory Policy"]
    end
    S1 --> A1 & T1 --> SCORE --> E1 --> REG["v_model_leaderboard"]

Act 1 · Cloud Bootstrap & Deployment Resolution¶

On a managed cloud notebook (Colab Enterprise, Vertex AI Workbench) the bootstrap cell below clones or updates the repository and installs the locked dependency set (uv.lock) into an isolated .colab-deps prefix so import scale_forecasting resolves against the exact package versions every cloud worker uses. In a local clone with .venv, this cell is a fast no-op.

In [1]:

# Cloud bootstrap: clone + install the LOCKED dependency set (uv.lock) so
# `import scale_forecasting` resolves against the exact versions every other surface runs.
import importlib.util
import os
import shutil
import subprocess
import sys

REPO_URL = os.environ.get("SF_REPO_URL", "https://github.com/statmike/scale-forecasting.git")
REPO_DIR = os.environ.get("SF_REPO_DIR", "scale-forecasting")
EXTRAS = ['gcp']

if importlib.util.find_spec("scale_forecasting") is None:
    if not os.path.isdir(REPO_DIR):
        subprocess.run(["git", "clone", "--depth", "1", REPO_URL, REPO_DIR], check=True)
    uv = shutil.which("uv")
    if uv is None:
        subprocess.run([sys.executable, "-m", "pip", "install", "-q", "uv"], check=True)
        uv = shutil.which("uv") or "uv"
    reqs = os.path.abspath(os.path.join(REPO_DIR, "colab-requirements.txt"))
    extra_flags = [f for e in EXTRAS for f in ("--extra", e)]
    subprocess.run(
        [uv, "export", "--frozen", "--no-emit-project", "--no-hashes", "--no-dev", *extra_flags, "-o", reqs],
        cwd=REPO_DIR,
        check=True,
    )
    target = os.path.abspath(os.path.join(REPO_DIR, ".colab-deps"))
    subprocess.run(
        [uv, "pip", "install", "--python", sys.executable, "--target", target, "-r", reqs],
        check=True,
    )
    sys.path.insert(0, target)
    sys.path.insert(0, os.path.join(REPO_DIR, "src"))
    print("installed scale_forecasting from", REPO_DIR)
else:
    print("scale_forecasting already importable — skipping bootstrap")

# Cloud bootstrap: clone + install the LOCKED dependency set (uv.lock) so # `import scale_forecasting` resolves against the exact versions every other surface runs. import importlib.util import os import shutil import subprocess import sys REPO_URL = os.environ.get("SF_REPO_URL", "https://github.com/statmike/scale-forecasting.git") REPO_DIR = os.environ.get("SF_REPO_DIR", "scale-forecasting") EXTRAS = ['gcp'] if importlib.util.find_spec("scale_forecasting") is None: if not os.path.isdir(REPO_DIR): subprocess.run(["git", "clone", "--depth", "1", REPO_URL, REPO_DIR], check=True) uv = shutil.which("uv") if uv is None: subprocess.run([sys.executable, "-m", "pip", "install", "-q", "uv"], check=True) uv = shutil.which("uv") or "uv" reqs = os.path.abspath(os.path.join(REPO_DIR, "colab-requirements.txt")) extra_flags = [f for e in EXTRAS for f in ("--extra", e)] subprocess.run( [uv, "export", "--frozen", "--no-emit-project", "--no-hashes", "--no-dev", *extra_flags, "-o", reqs], cwd=REPO_DIR, check=True, ) target = os.path.abspath(os.path.join(REPO_DIR, ".colab-deps")) subprocess.run( [uv, "pip", "install", "--python", sys.executable, "--target", target, "-r", reqs], check=True, ) sys.path.insert(0, target) sys.path.insert(0, os.path.join(REPO_DIR, "src")) print("installed scale_forecasting from", REPO_DIR) else: print("scale_forecasting already importable — skipping bootstrap")

scale_forecasting already importable — skipping bootstrap

Resolve the Live GCP Deployment (Settings.resolve())¶

Settings.resolve() reads the SF_* environment variables (SF_PROJECT_ID, SF_REGION, SF_DATASET_ID, SF_WAREHOUSE_URI, SF_CONNECTION, SF_COMPUTE_SA) pre-populated by Terraform in the sf-main runtime template (or your local shell).

In [2]:

from google.cloud import bigquery
from scale_forecasting.settings import Settings

settings = Settings.resolve()
client = bigquery.Client(project=settings.project_id)
DATASET = settings.dataset_ref
print(f"Deployment: {DATASET} | Region: {settings.region} | Bucket: {settings.warehouse_uri}")

from google.cloud import bigquery from scale_forecasting.settings import Settings settings = Settings.resolve() client = bigquery.Client(project=settings.project_id) DATASET = settings.dataset_ref print(f"Deployment: {DATASET} | Region: {settings.region} | Bucket: {settings.warehouse_uri}")

Deployment: statmike-scale-forecasting.scale_forecasting | Region: us-central1 | Bucket: gs://statmike-scale-forecasting-warehouse/warehouse

Act 2 · Configure the Native SQL Run & Explain the Execution Plan¶

We configure both BigQuery-native model families with univariate and covariate-enabled variants (arima_plus, arima_plus_xreg, timesfm, timesfm@univariate), known future covariates (promo_flag, price_index), historical past covariates (temperature), a 2-fold rolling-origin backtest, and in-run ensembling (mean, inverse_error, nnls, champion). Calling forecaster.explain() validates the configuration and prints the execution plan before any query is submitted.

In [3]:

import time
from scale_forecasting import Forecaster, RunConfig

# === Parameters — edit me ===============================================
RUN_NAME = f"nb01 bq native {int(time.time())}"
SOURCE_TABLE = "source_series_covariates_native"  # flip to "source_series_covariates_iceberg" to read BigLake Iceberg
MODELS = ["arima_plus", "arima_plus_xreg", "timesfm", "timesfm@univariate"]
HORIZON = 28
SERIES_LIMIT = 20
N_FOLDS = 2
ENSEMBLE_STRATEGIES = ["mean", "inverse_error", "nnls", "champion"]
# ========================================================================

cfg = RunConfig(
    run_name=RUN_NAME,
    data={"source_table": SOURCE_TABLE, "horizon": HORIZON, "series_limit": SERIES_LIMIT},
    models=MODELS,
    features={
        "holidays": ["US"],
        "future_covariates": ["promo_flag", "price_index"],
        "past_covariates": ["temperature"],
    },
    backtest={"enabled": True, "n_folds": N_FOLDS, "horizon": HORIZON, "step": HORIZON},
    ensemble={"enabled": True, "strategies": ENSEMBLE_STRATEGIES},
)

forecaster = Forecaster(cfg, settings=settings)
print("Deterministic run_id:", forecaster.run_id)
display(forecaster.explain())  # pass validate_data=True to also run BigQuery pre-flight table checks

import time from scale_forecasting import Forecaster, RunConfig # === Parameters — edit me =============================================== RUN_NAME = f"nb01 bq native {int(time.time())}" SOURCE_TABLE = "source_series_covariates_native" # flip to "source_series_covariates_iceberg" to read BigLake Iceberg MODELS = ["arima_plus", "arima_plus_xreg", "timesfm", "timesfm@univariate"] HORIZON = 28 SERIES_LIMIT = 20 N_FOLDS = 2 ENSEMBLE_STRATEGIES = ["mean", "inverse_error", "nnls", "champion"] # ======================================================================== cfg = RunConfig( run_name=RUN_NAME, data={"source_table": SOURCE_TABLE, "horizon": HORIZON, "series_limit": SERIES_LIMIT}, models=MODELS, features={ "holidays": ["US"], "future_covariates": ["promo_flag", "price_index"], "past_covariates": ["temperature"], }, backtest={"enabled": True, "n_folds": N_FOLDS, "horizon": HORIZON, "step": HORIZON}, ensemble={"enabled": True, "strategies": ENSEMBLE_STRATEGIES}, ) forecaster = Forecaster(cfg, settings=settings) print("Deterministic run_id:", forecaster.run_id) display(forecaster.explain()) # pass validate_data=True to also run BigQuery pre-flight table checks

Deterministic run_id: nb01-bq-native-1792004174-754186bc0ccc

	run_id	family	job_key	runtime	machine_type	workers	hardware	gpu_type	models	training_modes	n_series	n_folds	expected_cells	covariates	depends_on
0	nb01-bq-native-1792004174-754186bc0ccc	native	sf-nb01-bq-native-1792004174-754186bc0ccc-nati...	bigquery	sql	1	sql	None	arima_plus, arima_plus@xreg, timesfm, timesfm@...	arima_plus:local, arima_plus@xreg:local, times...	20	2	80	future(2), past(1)
1	nb01-bq-native-1792004174-754186bc0ccc	ensemble	sf-nb01-bq-native-1792004174-754186bc0ccc-ense...	bigquery	sql	1	sql	None	ensemble_mean, ensemble_inverse_error, ensembl...	barrier	20	2	80	future(2), past(1)	sf-nb01-bq-native-1792004174-754186bc0ccc-nati...

Act 3 · Launch & Monitor Live (forecaster.run_live())¶

forecaster.run_live() launches forecaster.run() on a background daemon thread, renders a live in-place progress dashboard via forecaster.monitor(), and returns the completed RunResult.

In [4]:

result = forecaster.run_live(poll_s=10.0, probe=True)
print(f"Run {result.run_id} finished with status={result.status} in {result.runtime_seconds:.1f}s")

result = forecaster.run_live(poll_s=10.0, probe=True) print(f"Run {result.run_id} finished with status={result.status} in {result.runtime_seconds:.1f}s")

Run nb01-bq-native-1792004174-754186bc0ccc finished with status=COMPLETED in 305.1s

Act 4 · Retrieve & Review Results (Leaderboard, Distributions & Forecast Fan Charts)¶

With the run complete, we inspect:

1. forecaster.leaderboard_df(): Ranked model + ensemble table from v_model_leaderboard.
2. plot_leaderboard() & plot_metric_distribution(): Visual ranking and per-series error quantiles (p10/p50/p90).
3. forecaster.plot_forecasts(): Historical series + backtest OOF overlay + 80% prediction intervals.
4. forecaster.trace() & plot_trace(): Execution timeline across families.

In [5]:

import matplotlib.pyplot as plt
from scale_forecasting import plot_leaderboard, plot_metric_distribution, plot_trace

lb_df = forecaster.leaderboard_df()
display(lb_df)

review = forecaster.review_run()
plot_leaderboard(review)
plt.show()

plot_metric_distribution(review)
plt.show()

import matplotlib.pyplot as plt from scale_forecasting import plot_leaderboard, plot_metric_distribution, plot_trace lb_df = forecaster.leaderboard_df() display(lb_df) review = forecaster.review_run() plot_leaderboard(review) plt.show() plot_metric_distribution(review) plt.show()

	rank	model_type	base_model	family	scope_capability	training_mode	cov_profile	is_univariate	covariates_used	model_spec_hash	...	mean_mae	mean_rmse	mean_coverage	mean_interval_score	mean_fit_seconds	median_fit_seconds	no_artifact_rate	n_predictions	lift_vs_best_base	lift_pct
0	1	ensemble_champion	ensemble_champion	ensemble	None	None	None	None	None	None	...	11.115575	17.742172	0.683036	65.583338	None	None	1.0	560	0.001996	0.005577
1	2	timesfm	timesfm	native	local_only	local	cov_dynamic	False	promo_flag,price_index,temperature	30a0ea9f89beebbe	...	11.200610	18.176520	0.799107	57.600348	None	None	1.0	560	NaN	NaN
2	3	timesfm@univariate	timesfm	native	local_only	local	cov_none	True		d80ff313a6de8f0c	...	11.462633	18.300876	0.770536	58.269985	None	None	1.0	560	NaN	NaN
3	4	ensemble_nnls	ensemble_nnls	ensemble	None	None	None	None	None	None	...	11.139548	17.834735	0.669643	67.006151	None	None	0.0	560	-0.012079	-0.033755
4	5	ensemble_inverse_error	ensemble_inverse_error	ensemble	None	None	None	None	None	None	...	11.080968	17.676197	0.668750	66.838714	None	None	1.0	560	-0.018985	-0.053054
5	6	ensemble_mean	ensemble_mean	ensemble	None	None	None	None	None	None	...	11.161460	17.772690	0.670536	67.468913	None	None	1.0	560	-0.022690	-0.063408
6	7	arima_plus@xreg	arima_plus	native	local_only	local	cov_future_only	False	promo_flag,price_index	8fcf5c3882890e5d	...	11.818180	18.302014	0.605357	63.348527	None	None	1.0	560	NaN	NaN
7	8	arima_plus	arima_plus	native	local_only	local	cov_none	True		a773fa5348c971a8	...	12.342413	18.937531	0.591964	68.670043	None	None	1.0	560	NaN	NaN

8 rows × 34 columns

In [6]:

# Inspect forecast fan chart + backtest OOF overlay on a single series
preds_df = forecaster.predictions_df(include_history=True, include_oof=True, series_limit=1)
sample_ts = str(preds_df["ts_id"].iloc[0])
print(f"Plotting forecasts for series: {sample_ts}")
forecaster.plot_forecasts(ts_id=sample_ts, history_tail=120)
plt.show()

# Inspect family execution trace
trace_df = forecaster.trace()
display(trace_df)

# Inspect forecast fan chart + backtest OOF overlay on a single series preds_df = forecaster.predictions_df(include_history=True, include_oof=True, series_limit=1) sample_ts = str(preds_df["ts_id"].iloc[0]) print(f"Plotting forecasts for series: {sample_ts}") forecaster.plot_forecasts(ts_id=sample_ts, history_tail=120) plt.show() # Inspect family execution trace trace_df = forecaster.trace() display(trace_df)

Plotting forecasts for series: s_000000

	kind	lane	label	start	end	duration_s	status	runtime	model_type	ts_id
0	job	ensemble	ensemble	2026-10-14 19:00:41.121579+00:00	2026-10-14 19:01:04.830874+00:00	23.709295	COMPLETED	bigquery	None	None
1	job	native	native	2026-10-14 18:56:47.209878+00:00	2026-10-14 19:00:35.098418+00:00	227.888540	COMPLETED	bigquery	None	None

Act 5 · Deep Dive: 21-Metric Panel, Covariate FVA, ML.EXPLAIN_FORECAST Attributions, Anomalies, Change Points & Inventory Policy¶

1. 21-Metric Panel & Covariate FVA (leaderboard_df(all_metrics=True) & fva_df()): Quantify whether adding exogenous covariates improved arima_plus@xreg over univariate arima_plus and timesfm (TimesFM 3.0) over timesfm@univariate (TimesFM 2.5) with Diebold–Mariano $p$-values.
2. Two-Tier ML.EXPLAIN_FORECAST Attributions (attributions_df() & plot_attributions()): Inspect Tier 1 global feature importance and Tier 2 per-step forecast attributions (trend, seasonal, holiday_effect, promo_flag, price_index) from ARIMA_PLUS_XREG.
3. Conformal Interval Calibration & Forecast Decomposition (plot_calibration() & plot_forecast_explanation()): Compare empirical 80% interval coverage across lead steps $h = 1 \dots H$ and decompose series trajectories.
4. Universal Anomaly & Structural Change-Point Detection (detect_anomalies_df() & detect_change_points_df()): Run BigQuery's native ML.DETECT_ANOMALIES / ML.DETECT_CHANGE_POINTS alongside universal OOF residual segmentation.
5. Operational Inventory & Replenishment Policy (inventory_policy_df() & plot_inventory_policy()): Translate calibrated horizon forecasts and backtest residual autocorrelation into per-SKU Reorder Points (ROP), Safety Stock (SS), and recommended order quantities.

In [7]:

full_lb = forecaster.leaderboard_df(all_metrics=True)
display(full_lb)

print("--- Forecast Value Added (Total & Covariate FVA with Diebold-Mariano p-values) ---")
display(forecaster.fva_df())

print("--- Two-Tier ML.EXPLAIN_FORECAST Attributions (arima_plus@xreg) ---")
attr_global = forecaster.attributions_df(model_type="arima_plus@xreg", level="global")
display(attr_global.head(10))
forecaster.plot_attributions(
    model_type="arima_plus@xreg",
    level="global",
    title="Tier 1 Global ML.EXPLAIN_FORECAST Attributions (arima_plus@xreg)",
)
plt.show()
forecaster.plot_attributions(
    model_type="arima_plus@xreg",
    level="local",
    ts_id=sample_ts,
    title=f"Tier 2 Per-Step ML.EXPLAIN_FORECAST Attributions ({sample_ts})",
)
plt.show()

display(forecaster.calibration_df(by="horizon").head(10))
forecaster.plot_calibration()
plt.show()

forecaster.plot_forecast_explanation(ts_id=sample_ts, history_tail=120)
plt.show()

print("--- Anomaly & Structural Change-Point Detection ---")
anom_df = forecaster.detect_anomalies_df(model_name="arima_plus@xreg", anomalies_only=True)
display(anom_df.head(10))
cp_df = forecaster.detect_change_points_df()
display(cp_df.head(10))

print("--- Operational Inventory & Replenishment Policy (Continuous Review (s, S) @ 95% CSL) ---")
inv_df = forecaster.inventory_policy_df(
    model_name="ensemble_nnls",
    lead_time=7.0,
    service_level=0.95,
    on_hand=150.0,
    holding_cost=1.0,
    stockout_cost=9.0,
)
display(
    inv_df[
        [
            "ts_id",
            "model_type",
            "lead_time_demand",
            "error_autocorr_rho",
            "safety_stock",
            "reorder_point",
            "recommended_order_qty",
            "expected_total_cost",
        ]
    ].head(10)
)
forecaster.plot_inventory_policy(
    model_name="ensemble_nnls",
    lead_time=7.0,
    service_level=0.95,
    on_hand=150.0,
    max_rows=10,
)
plt.show()

full_lb = forecaster.leaderboard_df(all_metrics=True) display(full_lb) print("--- Forecast Value Added (Total & Covariate FVA with Diebold-Mariano p-values) ---") display(forecaster.fva_df()) print("--- Two-Tier ML.EXPLAIN_FORECAST Attributions (arima_plus@xreg) ---") attr_global = forecaster.attributions_df(model_type="arima_plus@xreg", level="global") display(attr_global.head(10)) forecaster.plot_attributions( model_type="arima_plus@xreg", level="global", title="Tier 1 Global ML.EXPLAIN_FORECAST Attributions (arima_plus@xreg)", ) plt.show() forecaster.plot_attributions( model_type="arima_plus@xreg", level="local", ts_id=sample_ts, title=f"Tier 2 Per-Step ML.EXPLAIN_FORECAST Attributions ({sample_ts})", ) plt.show() display(forecaster.calibration_df(by="horizon").head(10)) forecaster.plot_calibration() plt.show() forecaster.plot_forecast_explanation(ts_id=sample_ts, history_tail=120) plt.show() print("--- Anomaly & Structural Change-Point Detection ---") anom_df = forecaster.detect_anomalies_df(model_name="arima_plus@xreg", anomalies_only=True) display(anom_df.head(10)) cp_df = forecaster.detect_change_points_df() display(cp_df.head(10)) print("--- Operational Inventory & Replenishment Policy (Continuous Review (s, S) @ 95% CSL) ---") inv_df = forecaster.inventory_policy_df( model_name="ensemble_nnls", lead_time=7.0, service_level=0.95, on_hand=150.0, holding_cost=1.0, stockout_cost=9.0, ) display( inv_df[ [ "ts_id", "model_type", "lead_time_demand", "error_autocorr_rho", "safety_stock", "reorder_point", "recommended_order_qty", "expected_total_cost", ] ].head(10) ) forecaster.plot_inventory_policy( model_name="ensemble_nnls", lead_time=7.0, service_level=0.95, on_hand=150.0, max_rows=10, ) plt.show()

	rank	model_type	base_model	family	scope_capability	training_mode	cov_profile	is_univariate	covariates_used	model_spec_hash	...	mean_rmsle	p50_rmsle	mean_msse	p50_msse	mean_msis	p50_msis	mean_r2	p50_r2	mean_cv	p50_cv
0	1	ensemble_champion	ensemble_champion	ensemble	None	None	None	None	None	None	...	0.533580	0.318237	0.589100	0.581731	3.823628	3.111291	0.171807	0.032398	0.578251	0.403684
1	2	timesfm	timesfm	native	local_only	local	cov_dynamic	False	promo_flag,price_index,temperature	30a0ea9f89beebbe	...	0.539940	0.318237	0.595496	0.586444	3.409656	2.950583	0.166610	0.040350	0.582706	0.403684
2	3	timesfm@univariate	timesfm	native	local_only	local	cov_none	True		d80ff313a6de8f0c	...	0.541491	0.313700	0.610161	0.581731	3.454687	3.067895	0.147739	0.027918	0.583899	0.434743
3	4	ensemble_nnls	ensemble_nnls	ensemble	None	None	None	None	None	None	...	0.476238	0.312247	0.573389	0.574712	3.819906	3.133271	0.206975	0.031070	0.556651	0.429548
4	5	ensemble_inverse_error	ensemble_inverse_error	ensemble	None	None	None	None	None	None	...	0.472225	0.313618	0.565727	0.523033	3.733494	3.118753	0.227468	0.039803	0.542148	0.423129
5	6	ensemble_mean	ensemble_mean	ensemble	None	None	None	None	None	None	...	0.475904	0.313609	0.575096	0.504452	3.797129	3.137020	0.219137	-0.010487	0.540743	0.426203
6	7	arima_plus@xreg	arima_plus	native	local_only	local	cov_future_only	False	promo_flag,price_index	8fcf5c3882890e5d	...	0.503575	0.320699	0.616356	0.463883	3.928701	3.501738	0.174417	-0.005486	0.534234	0.462074
7	8	arima_plus	arima_plus	native	local_only	local	cov_none	True		a773fa5348c971a8	...	0.515521	0.317803	0.688080	0.462272	4.116204	3.579147	0.107344	-0.007644	0.544433	0.468773

8 rows × 68 columns

--- Forecast Value Added (Total & Covariate FVA with Diebold-Mariano p-values) ---

	fva_stage_rank	fva_stage	model_type	base_model	demand_class	n_series	n_points	metric	mean_metric	pooled_wape	...	univariate_baseline_model	covariate_fva_abs	covariate_fva_pct	covariate_p_value_dm	covariate_significant_at_05	local_baseline_model	pooling_fva_abs	pooling_fva_pct	pooling_p_value_dm	pooling_significant_at_05
0	1	Baseline	arima_plus	arima_plus	all	20	1120	wape	0.408116	0.139355	...	None	NaN	NaN	NaN	False	None	NaN	NaN	NaN	False
1	2	Statistical / SQL	timesfm	timesfm	all	20	1120	wape	0.348967	0.126463	...	timesfm@univariate	0.005923	0.016691	0.001100	True	None	NaN	NaN	NaN	False
2	2	Statistical / SQL	timesfm@univariate	timesfm	all	20	1120	wape	0.354890	0.129422	...	None	NaN	NaN	NaN	False	None	NaN	NaN	NaN	False
3	2	Statistical / SQL	arima_plus@xreg	arima_plus	all	20	1120	wape	0.399999	0.133436	...	arima_plus	0.008117	0.019889	0.243926	False	None	NaN	NaN	NaN	False
4	6	Ensemble	ensemble_champion	ensemble_champion	all	20	1120	wape	0.348443	0.125503	...	None	NaN	NaN	NaN	False	None	NaN	NaN	NaN	False
5	6	Ensemble	ensemble_nnls	ensemble_nnls	all	20	1120	wape	0.361538	0.125774	...	None	NaN	NaN	NaN	False	None	NaN	NaN	NaN	False
6	6	Ensemble	ensemble_inverse_error	ensemble_inverse_error	all	20	1120	wape	0.367915	0.125113	...	None	NaN	NaN	NaN	False	None	NaN	NaN	NaN	False
7	6	Ensemble	ensemble_mean	ensemble_mean	all	20	1120	wape	0.371468	0.126021	...	None	NaN	NaN	NaN	False	None	NaN	NaN	NaN	False

8 rows × 30 columns

--- Two-Tier ML.EXPLAIN_FORECAST Attributions (arima_plus@xreg) ---

	ts_id	model_type	compute_engine	feature	importance
0	s_000000	arima_plus@xreg	bigquery	price_index	0.658020
1	s_000000	arima_plus@xreg	bigquery	trend	0.184002
2	s_000000	arima_plus@xreg	bigquery	seasonal	0.135954
3	s_000000	arima_plus@xreg	bigquery	holiday_effect	0.011431
4	s_000000	arima_plus@xreg	bigquery	promo_flag	0.010593
5	s_000001	arima_plus@xreg	bigquery	price_index	0.639330
6	s_000001	arima_plus@xreg	bigquery	trend	0.302129
7	s_000001	arima_plus@xreg	bigquery	holiday_effect	0.057363
8	s_000001	arima_plus@xreg	bigquery	promo_flag	0.001178
9	s_000002	arima_plus@xreg	bigquery	trend	0.799161

	model_type	horizon_step	n	coverage	nominal_coverage	coverage_error	mean_width
0	arima_plus	1	40	0.575	0.8	-0.225	17.888902
1	arima_plus	2	40	0.625	0.8	-0.175	20.828999
2	arima_plus	3	40	0.650	0.8	-0.150	22.332859
3	arima_plus	4	40	0.550	0.8	-0.250	23.267542
4	arima_plus	5	40	0.650	0.8	-0.150	23.901164
5	arima_plus	6	40	0.650	0.8	-0.150	24.352030
6	arima_plus	7	40	0.500	0.8	-0.300	24.682402
7	arima_plus	8	40	0.725	0.8	-0.075	24.929076
8	arima_plus	9	40	0.575	0.8	-0.225	25.115641
9	arima_plus	10	40	0.700	0.8	-0.100	25.258098

--- Anomaly & Structural Change-Point Detection ---

	ts_id	model_type	ds	y	yhat	lower_bound	upper_bound	anomaly_probability	is_anomaly
0	s_000000	arima_plus@xreg	2021-01-02	227.510	245.665425	233.602866	257.727984	0.996303	True
1	s_000000	arima_plus@xreg	2021-01-03	231.420	212.685609	200.623050	224.748168	0.997171	True
2	s_000000	arima_plus@xreg	2021-02-27	232.967	248.857665	236.795106	260.920224	0.989714	True
3	s_000000	arima_plus@xreg	2021-03-17	191.745	206.155192	194.092633	218.217751	0.980512	True
4	s_000000	arima_plus@xreg	2021-03-19	265.128	252.546530	240.483971	264.609089	0.959044	True
5	s_000000	arima_plus@xreg	2021-03-26	263.737	280.605971	268.543412	292.668530	0.993353	True
6	s_000000	arima_plus@xreg	2021-03-27	288.160	303.659339	291.596780	315.721898	0.987788	True
7	s_000000	arima_plus@xreg	2021-04-24	292.134	304.928519	292.865960	316.991078	0.962323	True
8	s_000000	arima_plus@xreg	2021-05-19	152.588	139.181926	127.119367	151.244485	0.970488	True
9	s_000000	arima_plus@xreg	2021-06-17	176.599	162.310765	150.248206	174.373324	0.979487	True

	ts_id	model_type	begin_date	end_date	window_count	window_avg	window_min	window_max	window_stddev	gap_edge_missing_steps	is_gap_edge	status
0	s_000000	raw	2021-01-26	2021-02-07	13	188.502769	134.928	255.680	41.357512	0	False
1	s_000007	raw	2023-08-29	2023-09-04	7	153.746143	120.515	196.677	29.086596	0	False
2	s_000008	raw	2021-06-06	2021-06-25	20	49.747750	31.525	79.907	14.088549	0	False
3	s_000008	raw	2022-06-05	2022-06-28	24	74.430000	31.660	256.688	56.878722	0	False
4	s_000008	raw	2023-06-07	2023-06-25	19	50.390000	27.210	74.882	12.911749	0	False
5	s_000008	raw	2024-06-08	2024-06-26	19	50.859368	30.474	66.456	10.958760	0	False
6	s_000010	raw	2024-11-29	2024-12-09	11	134.021909	85.212	175.438	31.012458	0	False
7	s_000011	raw	2024-11-25	2024-12-09	15	4.795200	0.000	17.004	6.464117	0	False
8	s_000013	raw	2021-01-22	2021-02-05	15	73.517000	62.727	86.370	6.796223	0	False
9	s_000013	raw	2023-12-23	2024-01-11	20	93.137300	53.347	401.212	86.065814	0	False

--- Operational Inventory & Replenishment Policy (Continuous Review (s, S) @ 95% CSL) ---

	ts_id	model_type	lead_time_demand	error_autocorr_rho	safety_stock	reorder_point	recommended_order_qty	expected_total_cost
0	s_000000	ensemble_nnls	1590.134575	0.121169	87.766272	1677.900847	1527.900847	98.914359
1	s_000001	ensemble_nnls	5.280125	0.135588	26.481880	31.762006	0.000000	29.845613
2	s_000002	ensemble_nnls	1202.268299	0.597727	114.312882	1316.581181	1166.581181	128.832923
3	s_000003	ensemble_nnls	541.126591	0.411614	55.668980	596.795571	446.795571	62.740064
4	s_000004	ensemble_nnls	381.521922	0.676639	447.049843	828.571765	678.571765	503.834191
5	s_000005	ensemble_nnls	1659.825999	0.216267	44.168800	1703.994799	1553.994799	49.779128
6	s_000006	ensemble_nnls	3.510889	-0.031338	10.788479	14.299368	0.000000	12.158833
7	s_000007	ensemble_nnls	1687.643258	0.318294	131.270191	1818.913449	1668.913449	147.944153
8	s_000008	ensemble_nnls	370.989202	-0.053908	323.817477	694.806679	544.806679	364.948829
9	s_000009	ensemble_nnls	470.091226	0.806596	782.515153	1252.606379	1102.606379	881.910362

---

Where next¶

• Next in Track 2 (Cloud Runtimes): 02_vertex_and_gce_vms.ipynb — run Python models on self-deleting Compute Engine VMs (gce), serverless Vertex AI CustomJob (vertex), and GKE Indexed Jobs (gke).
• Compare runtimes: Choosing a runtime · Cost estimates & controls
• Documentation: Runtimes reference · Output schemas & views