anofox-forecast-backtest

Backtesting, cross-validation, evaluation metrics, and conformal prediction intervals for the anofox_forecast DuckDB extension. Use when evaluating forecast accuracy, comparing models with time-series

SKILL.md

GitHub

Anofox Forecast — Backtesting, CV, Metrics & Conformal Cheat Sheet

Extension: anofox_forecast v2026.09.26 (Rust crate anofox-forecast v0.16.2) | DuckDB: v1.4.5 LTS / v1.5.4+ | Dual naming: ts_* and anofox_fcst_ts_*

Time-series-aware cross-validation, error metrics, and distribution-free intervals.

Critical gotchas

  • ts_backtest_auto_by was REMOVED, along with the orphaned internal _ts_backtest_native table function that backed it (Phase 12 D-07). Use the two-step CV workflow (ts_cv_folds_by → ts_cv_forecast_by) instead.
  • Metric _by table macros are deprecated. They're ~2400× slower than the scalar + GROUP BY pattern and don't parallelise. Use scalars. The full deprecated family: ts_mae_by, ts_rmse_by, ts_mse_by, ts_mape_by, ts_smape_by, ts_bias_by, ts_mase_by, ts_rmae_by, ts_r2_by, ts_coverage_by, ts_quantile_loss_by, ts_interval_width_by. Each maps to a scalar (ts_mae, ts_rmse, … below) called under GROUP BY. Two helpers are exceptions: ts_interval_width_by is a table macro (results, group_col, lower_col, upper_col) and ts_mean_interval_width(DOUBLE[] lower, DOUBLE[] upper) → DOUBLE is a scalar — still prefer the scalar + GROUP BY pattern over the _by table macro.
  • Always ORDER BY inside LIST() for temporal correctness: LIST(y ORDER BY ds), not LIST(y).
  • ts_cv_forecast_by output renames the target column to y (canonical). Don't try to access the original name.
  • Folds must be pre-computed before forecasting. Passing raw data to ts_cv_forecast_by throws a clear error.
  • Metrics use pairwise-complete NULL handling (v2026.10.16, Phase 12). ts_mae/ts_mse/ts_rmse/ts_mape/ts_smape/ts_r2/ts_bias/ts_rmae/ts_quantile_loss/ts_mqloss/ts_coverage drop a pair (or row, for 3+ input metrics) the moment ANY input is NULL at that position, computed in Rust — not spliced independently per array in C++. No complete pair/row → NULL, not an error. ts_mape additionally returns NULL for the whole call if any pairwise-complete actual is exactly 0.0 (D-07) — use ts_smape/ts_mase for intermittent demand. ts_mase's insample argument has its OWN policy (never pairwise-complete-filtered — see below), not this one.

The CV two-step workflow (standard)

-- Setup: data(unique_id, ds, y), 2 series x 130 daily rows (shared by every example
-- below in this file).
CREATE OR REPLACE TABLE data AS
SELECT
    unique_id,
    (DATE '2020-01-01' + INTERVAL (i) DAY) AS ds,
    100.0 + unique_id * 10.0 + 0.3 * i + 5.0 * sin(2 * pi() * i / 7.0) AS y
FROM range(0, 130) t(i)
CROSS JOIN (SELECT unnest([1, 2]) AS unique_id);
-- Step 1: Create fold table (train/test rows with actual dates)
CREATE OR REPLACE TABLE cv_folds AS
SELECT * FROM ts_cv_folds_by('data', unique_id, ds, y,
    3,                  -- n_folds
    12,                 -- horizon per fold
    MAP{});             -- optional params

-- Step 2: Forecast per fold's train set, predict its test set
CREATE OR REPLACE TABLE cv_forecasts AS
SELECT * FROM ts_cv_forecast_by('cv_folds', unique_id, ds, y,
    'AutoETS',
    MAP{'seasonal_period': '12'});

-- Step 3: Compute per-series / per-fold metrics
SELECT unique_id, fold_id,
    ts_rmse(LIST(y ORDER BY ds), LIST(yhat ORDER BY ds)) AS rmse,
    ts_mae(LIST(y ORDER BY ds), LIST(yhat ORDER BY ds))  AS mae
FROM cv_forecasts
GROUP BY unique_id, fold_id;

ts_cv_folds_by

ts_cv_folds_by(source VARCHAR, group_col COLUMN, date_col COLUMN, target_col COLUMN,
               n_folds BIGINT, horizon BIGINT, params MAP/STRUCT) → TABLE

Input must be pre-cleaned (no gaps, consistent frequency). Uses position-based indexing so works with all frequencies (including calendar-based monthly / quarterly / yearly).

Params:

Key Default Description
gap 0 Periods between train end and test start
embargo 0 Periods excluded from training after previous test
window_type 'expanding' 'expanding', 'fixed', or 'sliding'
min_train_size 1 Min training size (fixed / sliding only)
initial_train_size auto Periods before first fold
skip_length horizon Periods between folds (1 = dense overlap)
clip_horizon false Allow partial test windows near series end

Output: 5 columns — <group_col>, <date_col>, <target_col>, fold_id BIGINT, split VARCHAR ('train' / 'test'). Features NOT passed through — use ts_cv_hydrate_by to join extra columns.

ts_cv_split_by / ts_cv_split_folds_by / ts_cv_split_index_by

Alternate fold-creation entry points for custom cutoffs, pre-computed fold boundaries, or memory-efficient index-only splits. See docs/api/08-cross-validation.md for their param signatures.

ts_cv_forecast_by

ts_cv_forecast_by(cv_folds_table VARCHAR, group_col COLUMN, date_col COLUMN, target_col COLUMN,
                  method VARCHAR, params MAP/STRUCT) → TABLE

Fits method on each fold's train partition, forecasts the test partition.

Output columns: fold_id, <group_col>, <date_col>, y (renamed from target_col), split, yhat, yhat_lower, yhat_upper, model_name.

Params: same shape as ts_forecast_by. All model strings work ('AutoETS', 'Laplace', 'HoltWinters', etc.). See anofox-forecast-models for the full catalogue.

ts_cv_hydrate_by — attach features to fold rows

If your model needs exogenous columns (features, calendar flags, promotions), the folds table lost them. Join them back:

-- Setup: features(unique_id, ds, promo, holiday), the raw source table that carries
-- the exogenous columns the folds table itself lost.
CREATE OR REPLACE TABLE features AS
SELECT unique_id, ds,
       (CAST(ds AS DATE) - DATE '2020-01-01') % 14 = 0 AS promo,
       (CAST(ds AS DATE) - DATE '2020-01-01') % 30 = 0 AS holiday
FROM data;

-- Signature is (cv_folds, source, group_col, date_col, unknown_features, params := MAP{})
CREATE OR REPLACE TABLE cv_folds_with_features AS
SELECT * FROM ts_cv_hydrate_by('cv_folds', 'features', unique_id, ds, ['promo', 'holiday']);

ts_check_leakage

Sanity-check that no test-window rows leaked into training. Signature is (source, is_test_col, params) — source is a table name (not a subquery string), and is_test_col is a boolean column on that table, not a list of positional group/date/fold columns:

CREATE OR REPLACE TABLE cv_folds_flagged AS
SELECT *, (split = 'test') AS is_test FROM cv_folds;

SELECT * FROM ts_check_leakage('cv_folds_flagged', is_test, MAP{});

Metrics — use scalar functions

Pattern: SELECT group, ts_<metric>(LIST(actual ORDER BY ds), LIST(pred ORDER BY ds)) FROM ... GROUP BY group;

Function Signature Description
ts_mae (DOUBLE[], DOUBLE[]) → DOUBLE Mean Absolute Error
ts_mse (DOUBLE[], DOUBLE[]) → DOUBLE Mean Squared Error
ts_rmse (DOUBLE[], DOUBLE[]) → DOUBLE Root MSE
ts_mape (DOUBLE[], DOUBLE[]) → DOUBLE Mean Absolute % Error — returns NULL if ANY pairwise-complete actual is exactly 0.0 (D-07; MAPE is undefined there) — use ts_smape/ts_mase for intermittent demand
ts_smape (DOUBLE[], DOUBLE[]) → DOUBLE Symmetric MAPE
ts_r2 (DOUBLE[], DOUBLE[]) → DOUBLE R²
ts_bias (DOUBLE[], DOUBLE[]) → DOUBLE Bias (mean error)
ts_mase (actual, forecast, insample, seasonal_period) → DOUBLE Hyndman-Koehler MASE (D-08): MAE(actual,forecast) / mean(|insample_t - insample_{t-seasonal_period}|). seasonal_period is a REQUIRED 4th positional arg (no := on scalar functions). insample lag pairs touching a missing value are skipped, never spliced. Scale numerically zero → NULL. The pre-v2026.10.16 3-arg (actual, forecast, baseline) form — a relative MAE, not MASE — is REMOVED; it errors naming ts_rmae. ts_mase_by(source, date_col, actual_col, forecast_col, seasonal_period := 1) needs source to include each group's HISTORY rows (forecast NULL, dated before the first forecast) to supply the in-sample scale — the old baseline-column 5th argument is also removed, naming ts_rmae_by.
ts_rmae (actual, pred1, pred2) → DOUBLE Relative MAE. < 1 = pred1 better
ts_coverage (actual, lower, upper) → DOUBLE Interval coverage rate
ts_quantile_loss (actual, forecast, q) → DOUBLE Quantile loss at level q
ts_mqloss (actual, quantiles[], levels[]) → DOUBLE Multi-quantile CRPS-adjacent

Model comparison across CV folds

-- Fit two models
CREATE OR REPLACE TABLE cv_naive   AS
SELECT * FROM ts_cv_forecast_by('cv_folds', unique_id, ds, y, 'Naive',   MAP{});
CREATE OR REPLACE TABLE cv_autoets AS
SELECT * FROM ts_cv_forecast_by('cv_folds', unique_id, ds, y, 'AutoETS', MAP{'seasonal_period': '7'});

-- Compare
SELECT 'Naive' AS model,
    ts_mae(LIST(y ORDER BY ds), LIST(yhat ORDER BY ds)) AS mae
FROM cv_naive
UNION ALL
SELECT 'AutoETS',
    ts_mae(LIST(y ORDER BY ds), LIST(yhat ORDER BY ds))
FROM cv_autoets
ORDER BY mae;

Conformal prediction — distribution-free intervals

Attach a distribution-free prediction interval to forecasts, calibrated from backtest residuals, horizon-ordered and calibrated per group by default (Phase 16). Coverage is guaranteed under exchangeability AND a finite-sample minimum calibration-set size for the requested alpha — see docs/api/11-conformal-prediction.md#coverage-guarantee-and-its-limits.

Three flavours

Approach Functions Use when
One-step ts_conformal_by You already have backtest results, want intervals now
Modular ts_conformal_calibrate + ts_conformal_apply_by Reuse calibration across multiple forecast rounds
Array-based ts_conformal_predict, ts_conformal_quantile, etc. Custom pipelines over LIST(...) arrays

ts_conformal_by (one-step)

ts_conformal_by(backtest_results VARCHAR, group_col COLUMN, date_col COLUMN,
                actual_col COLUMN, forecast_col COLUMN, point_forecast_col COLUMN,
                params := MAP{}, difficulty_col := NULL) → TABLE

Returns one row per (group_col, date_col). date_col is required and orders/pairs every row — forecasts are never value-sorted. Params:

Key Default Description
alpha 0.1 Miscoverage rate (0.1 = 90 %, 0.05 = 95 %)
method 'symmetric' 'symmetric' (alias 'split'), 'asymmetric' (skewed residuals), or 'adaptive' (needs difficulty_col)

coverage on the output is the EMPIRICAL fraction of calibration residuals covered, not simply 1 - alpha; a separate nominal_level column reports the requested level. calibration_insufficient = true (with NULL bounds/score) when the calibration set is below the alpha's minimum size (9 for alpha=0.1 symmetric, 19 asymmetric) — never clamped.

CREATE OR REPLACE TABLE conformal_bounds AS
SELECT * FROM ts_conformal_by('backtest', unique_id, ds, y, yhat, yhat, params := {'alpha': 0.1});

Modular (ts_conformal_calibrate → ts_conformal_apply_by)

-- Compute calibration once -- one row PER GROUP by default (D-03); pool via
-- params := {'pooled': true}
CREATE OR REPLACE TABLE calib AS
SELECT * FROM ts_conformal_calibrate('backtest', unique_id, y, yhat, params := {'alpha': 0.1});

-- Apply to any future forecast -- `calib` is a calibration TABLE name (D-14), not a bare
-- score; each forecast group matches its own calibration row (falling back to a pooled
-- row with group_col IS NULL, if present)
SELECT * FROM ts_conformal_apply_by('new_forecasts', unique_id, ds, yhat, 'calib');

Array-based helpers

  • ts_conformal_predict, ts_conformal_predict_asymmetric, ts_conformal_predict_per_step
  • ts_conformal_quantile, ts_conformal_intervals
  • ts_conformal_learn, ts_conformal_apply — strategies are 'split'/'crossval' only; 'jackknife+' (and every alias) was removed in Phase 16 — it computed the identical quantile to 'split', with no actual leave-one-out refits behind the name
  • ts_conformal_coverage, ts_conformal_evaluate

Use these when composing custom pipelines over LIST(residual) arrays.

NULL policy (gotcha): residual pools drop NULL/NaN entries; forecast positions (ts_conformal_intervals/predict(_asymmetric)/apply/predict_per_step, both bootstrap functions) are position-preserving (NULL in → NULL at that position out, never a splice); ts_conformal_predict_per_step errors in Rust if the flat residual count isn't an exact multiple of horizon, or if point_forecasts length != horizon (try(...) to downgrade to NULL) — see docs/api/11-conformal-prediction.md#null-handling.

Bootstrap prediction intervals

Residual-bootstrap alternative to conformal — resample forecast residuals to build empirical intervals / quantiles around point forecasts. Both are scalar functions over DOUBLE[] arrays with a fixed seed for reproducibility. A right-skewed residual correctly puts the long tail on the UPPER side (Phase 16 fix, D-10).

ts_bootstrap_intervals — coverage-band intervals

ts_bootstrap_intervals(residuals DOUBLE[], forecasts DOUBLE[],
                       n_paths INTEGER, coverage DOUBLE, seed BIGINT)
    → STRUCT(point DOUBLE[], lower DOUBLE[], upper DOUBLE[], coverage DOUBLE)

n_paths must be >= 10 and n_paths * len(forecasts) <= 10,000,000; coverage must be in (0, 1) — each is a caller error naming the value, never silently adjusted.

SELECT ts_bootstrap_intervals(
    [0.2, -0.3, 0.1, -0.2, 0.15, -0.1, 0.05, -0.05],  -- residuals
    [10.0, 11, 10.5, 12, 11.5, 13, 12.5, 14],         -- point forecasts (7+ steps)
    1000, 0.90, 42) AS r;

ts_bootstrap_quantiles — arbitrary quantile levels

ts_bootstrap_quantiles(residuals DOUBLE[], forecasts DOUBLE[],
                       n_paths INTEGER, quantiles DOUBLE[], seed BIGINT)
    → STRUCT(point DOUBLE[], quantiles DOUBLE[], "values" DOUBLE[])

values is flattened quantiles × horizon (one block of horizon values per requested quantile level).

SELECT ts_bootstrap_quantiles(
    [0.2, -0.3, 0.1, -0.2, 0.15, -0.1, 0.05, -0.05],
    [10.0, 11, 10.5, 12, 11.5, 13, 12.5, 14],
    1000, [0.1, 0.5, 0.9], 42) AS r;

ts_estimate_backtest_memory

Pre-flight a CV run's memory footprint (scalar function: (n_series, n_dates, folds, horizon) → DOUBLE, estimated bytes):

SELECT ts_estimate_backtest_memory(1000, 365, 5, 14) AS estimated_bytes;

Full backtest → conformal pipeline

-- 1. CV folds
CREATE OR REPLACE TABLE folds AS
SELECT * FROM ts_cv_folds_by('clean', product_id, ds, y, 5, 14, MAP{});

-- 2. Model forecast per fold
CREATE OR REPLACE TABLE bt AS
SELECT * FROM ts_cv_forecast_by('folds', product_id, ds, y, 'AutoETS',
    MAP{'seasonal_period': '7'});

-- 3. Per-series metrics
SELECT product_id,
    ts_mae(LIST(y ORDER BY ds), LIST(yhat ORDER BY ds)) AS mae,
    ts_rmse(LIST(y ORDER BY ds), LIST(yhat ORDER BY ds)) AS rmse,
    ts_coverage(LIST(y ORDER BY ds), LIST(yhat_lower ORDER BY ds), LIST(yhat_upper ORDER BY ds)) AS cov
FROM bt GROUP BY product_id;

-- 4. Conformalise per product -> targets 90% nominal coverage (empirical coverage and
--    calibration_insufficient are reported alongside, see docs/api/11-conformal-prediction.md)
CREATE OR REPLACE TABLE calib AS
SELECT * FROM ts_conformal_calibrate('bt', product_id, y, yhat, params := {'alpha': 0.1});

-- 5. Apply calibration (a TABLE, not a bare score) to new forecasts
CREATE OR REPLACE TABLE new_fcst AS
SELECT * FROM ts_forecast_by('clean', product_id, ds, y, 'AutoETS', 14, '1d',
    MAP{'seasonal_period': '7'});

SELECT * FROM ts_conformal_apply_by('new_fcst', product_id, ds, yhat, 'calib');

See also: anofox-forecast-data-prep (clean input required by ts_cv_folds_by), anofox-forecast-detection (detect seasonal_period before backtesting), anofox-forecast-models (any model string usable in ts_cv_forecast_by).

Reference docs:

  • docs/api/08-cross-validation.md
  • docs/api/09-evaluation-metrics.md
  • docs/api/11-conformal-prediction.md