anofox-forecast-eda

Exploratory data analysis, data quality, and statistical diagnostics for the anofox_forecast DuckDB extension — 34 per-series statistics, data-quality scoring, quality-report summaries, 117 tsfresh-co

SKILL.md

GitHub

Anofox Forecast — EDA & Data Quality 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_*

Understand your data before modelling it: per-series statistics, data-quality scores, feature vectors for ML.

Statistics

ts_stats (table function) — 36 metrics per series

ts_stats(source VARCHAR, group_col COLUMN, date_col COLUMN, value_col COLUMN,
         frequency VARCHAR, seasonal_period := 0) → TABLE

Returns 36 columns: length, n_nulls, n_nan, n_zeros, n_positive, n_negative, n_unique_values, is_constant, n_zeros_start, n_zeros_end, plateau_size, plateau_size_nonzero, mean, median, std_dev, variance, min, max, range, sum, skewness, kurtosis, tail_index, bimodality_coef, trimmed_mean, coef_variation, q1, q3, iqr, autocorr_lag1, trend_strength, seasonality_strength, entropy, stability, plus date-derived expected_length, n_gaps, and n_gap_runs.

v2026.10.16: skewness/kurtosis are now textbook scipy bias=False G1/ excess-G2 (previously mixed a population moment with a sample std-dev); bimodality_coef uses the SAS small-sample formula; entropy is now Pincus (1991) approximate entropy (m=2, r=0.2*sample sd; NULL for an interior-missing or fewer-than-3-present series) — previously a 10-bin Shannon entropy; n_gaps is now the number of missing periods (previously gap occurrences) — the old occurrence count is the new n_gap_runs column. trend_strength/seasonality_strength are now STL-based (F_T/F_S, Wang/Hyndman) via a new seasonal_period parameter (see "Strength thresholds" below). is_constant (here and in ts_data_quality) is SCALE-RELATIVE (D-13, G25): max - min <= 1e-10 * max(1, max|x|) over present (non-NULL, non-NaN) values — NOT the same thing as n_unique_values == 1, which stays an exact bit-identity count. [1e12, 1e12 + 0.001] is now constant; [1.0, 1.0 + 1e-9] is not (the absolute-epsilon tolerance before this plan was the inverse on both).

-- Every stat for every series
SELECT * FROM ts_stats('sales', product_id, ds, y, '1d');

-- Quick sanity check: n_gaps = missing periods, n_gap_runs = gap occurrences
SELECT product_id, length, n_nulls, n_gaps, n_gap_runs, trend_strength, seasonality_strength
FROM ts_stats('sales', product_id, ds, y, '1d')
WHERE length < 30 OR n_gaps > 0
ORDER BY n_gaps DESC;

-- Explicit weekly period instead of auto-detection
SELECT product_id, trend_strength, seasonality_strength
FROM ts_stats('sales', product_id, ds, y, '1d', seasonal_period := 7);

ts_stats_agg (aggregate)

For custom GROUP BY shapes. Takes (date, value) — DO NOT wrap in LIST(...).

ts_stats_agg(date_col TIMESTAMP, value_col DOUBLE)
    → STRUCT(length UBIGINT, n_nulls UBIGINT, …, mean DOUBLE, …)
SELECT product_id,
       ts_stats_agg(ds, y).mean AS mean,
       ts_stats_agg(ds, y).seasonality_strength AS seas_str
FROM sales GROUP BY product_id;

ts_stats_by — alias for ts_stats

ts_stats_summary — aggregate stats across all groups

Summarise the per-group stats into panel-level statistics (mean, std, min, max of each metric). Signature is (stats_table) — a single argument naming an already-materialized ts_stats_by output table, not a raw source:

CREATE OR REPLACE TABLE sales_stats AS
SELECT * FROM ts_stats_by('sales', product_id, ds, y, '1d');

SELECT * FROM ts_stats_summary('sales_stats');

Data quality

ts_data_quality (table function) — per-series score card

ts_data_quality(source VARCHAR, unique_id_col COLUMN, date_col COLUMN, value_col COLUMN,
                n_short INTEGER, frequency VARCHAR) → TABLE

Returns unique_id (group column renamed — the extension normalises it to unique_id in this output, even if the input col was product_id), plus overall_score, structural_score, temporal_score, magnitude_score, behavioral_score, n_gaps, n_gap_runs, n_missing, is_constant, is_short.

n_short is the minimum "not short" present-value count (typical: 14 for daily, 12 for monthly) — it drives structural_score's length factor and the is_short flag. frequency ('1d', '1mo', ... or NULL/'' to auto-detect) is used, together with each series' real dates, for n_gaps/n_gap_runs/temporal_score:

  • n_gaps = missing periods (expected timestamps absent under the frequency) — same meaning as ts_stats's n_gaps.
  • n_gap_runs = gap occurrences (contiguous runs of missing periods).
  • temporal_score = clamp(1 - 5 * n_gaps / expected_length, 0, 1).
  • behavioral_score scores constancy, robust (Iglewicz-Hoaglin) outliers, and long zero-runs only — a smooth/predictable series carries no autocorrelation penalty.
SELECT * FROM ts_data_quality('sales', product_id, ds, y, 14, '1d');

-- Rank series by quality (note: output col is `unique_id`, not `product_id`)
SELECT unique_id, overall_score, structural_score, temporal_score, n_gaps, is_short
FROM ts_data_quality('sales', product_id, ds, y, 14, '1d')
ORDER BY overall_score;

ts_data_quality_by — table-macro variant

Grouped macro sibling of ts_data_quality — same per-series quality scoring by group. Signature ts_data_quality_by(source, unique_id_col, date_col, value_col, n_short, frequency).

ts_data_quality_agg (aggregate)

Takes (date, value), returns a STRUCT — no n_short/frequency arguments: n_short is always 30 and frequency always auto-detects from the group's own timestamps. Keeps NULL-valued rows (counted in n_missing); skips only NULL-timestamp rows.

SELECT product_id,
       ts_data_quality_agg(ds, y).overall_score AS q
FROM sales GROUP BY product_id;

ts_data_quality_summary — panel-level roll-up

Returns n_total, n_good/n_fair/n_poor (overall_score buckets: >=0.8, 0.5-0.8, <0.5), n_unassessed (groups with no usable dates, e.g. all-NULL date_col — excluded from the other buckets and from avg_score; D-10) and avg_score (mean over assessed groups only). Invariant: n_good + n_fair + n_poor + n_unassessed = n_total.

SELECT * FROM ts_data_quality_summary('sales', product_id, ds, y, 14);

ts_quality_report — human-readable report

ts_validate_timestamps_summary_by — per-group timestamp validation

Summarises timestamp regularity per group against an expected grid — flags gaps, duplicates, and irregular spacing before forecasting. Signature ts_validate_timestamps_summary_by(source, group_col, date_col, expected_timestamps).

Diagnostics & validation (v0.7.0)

Requires: json extension (same as detection). Enable auto-load once per session:

SET autoinstall_known_extensions = 1;
SET autoload_known_extensions = 1;

Seven functions in two groups. The combined-verdict functions (ts_stationarity, ts_residual_diagnostics) are the recommended entry points — they run the component tests internally and return a single adequacy flag.

Stationarity trio

Full detail in docs/api/10-diagnostics.md. ts_adf/ts_kpss/ts_stationarity (and their _by macros) take a regression := 'c' | 'ct' parameter (default 'c'): 'c' tests around a constant mean (statsmodels/tseries' default), 'ct' tests around a constant + linear time trend. ts_adf matches statsmodels.tsa.stattools.adfuller(autolag="AIC") exactly (AIC lag selection, MacKinnon finite-sample critical values/p-values); ts_kpss matches R tseries::kpss.test() for both 'c' (null = "Level") and 'ct' (null = "Trend", a trend KPSS implemented in this extension's Rust core since the upstream crate has no trend variant); KPSS p-values are bounded to [0.01, 0.10].

ts_adf / ts_adf_by — Dickey-Fuller unit-root test (statsmodels-exact, AIC lag selection)

H0: series has a unit root (non-stationary). Reject H0 (p < 0.05) → stationary.

-- Scalar form: takes LIST(value ORDER BY date) (sourced from stationarity.sql Section 2)
SELECT
    (adf).statistic       AS t_statistic,
    (adf).p_value         AS p_value,
    (adf).lags            AS lags_used,
    (adf).is_stationary   AS is_stationary,
    ROUND((adf).cv_5pct, 3) AS critical_value_5pct
FROM (
    SELECT ts_adf(LIST(y ORDER BY ds)) AS adf
    FROM sales_data
    WHERE product_id = 'mean_revert'
);

-- Grouped macro (sourced from stationarity.sql Section 3)
SELECT product_id,
       ROUND((adf).statistic, 4) AS t_statistic,
       (adf).p_value,
       (adf).lags,
       (adf).is_stationary,
       ROUND((adf).cv_1pct, 2) AS cv_1pct,
       ROUND((adf).cv_5pct, 2) AS cv_5pct,
       ROUND((adf).cv_10pct, 2) AS cv_10pct
FROM ts_adf_by('sales_data', product_id, ds, y)
ORDER BY product_id;

Optional second/third args to scalar form: ts_adf(LIST(y ORDER BY ds), max_lags INTEGER, regression VARCHAR) — max_lags := -1 = auto (default: ⌊(n−1)^(1/3)⌋, AIC-selected up to that bound); regression := 'c' (default) or 'ct'.

ts_kpss / ts_kpss_by — KPSS level/trend-stationarity test

H0: series IS level-stationary (opposite direction from ADF). Fail to reject (p > 0.05) → stationary.

-- Scalar form (sourced from stationarity.sql Section 5)
WITH s AS (SELECT i AS ds, sin(i/6.0) + (i%5)*0.01 AS y FROM range(1, 80) t(i))
SELECT (ts_kpss(LIST(y ORDER BY ds))).statistic     AS kpss_stat,
       (ts_kpss(LIST(y ORDER BY ds))).is_stationary AS is_stationary
FROM s;

-- Grouped macro (sourced from stationarity.sql Section 5)
SELECT product_id, (kpss).statistic, (kpss).is_stationary
FROM ts_kpss_by('sales_data', product_id, ds, y)
ORDER BY product_id;

ts_stationarity / ts_stationarity_by — combined four-way verdict (recommended entry point)

Runs ADF + KPSS internally; returns a combined verdict field plus the individual flags.

-- Grouped macro (sourced from stationarity.sql Section 6)
SELECT product_id,
       (stationarity).verdict,
       (stationarity).adf_is_stationary,
       (stationarity).kpss_is_stationary
FROM ts_stationarity_by('sales_data', product_id, ds, y)
ORDER BY product_id;

Verdict strings (the standard ADF+KPSS four-way mapping): 'stationary' (ADF rejects the unit root AND KPSS fails to reject level-stationarity), 'trend_stationary' (ADF does NOT reject but KPSS fails to reject — stationary only after removing a trend; use regression := 'ct' to see this), 'non_stationary' (ADF does NOT reject AND KPSS rejects), 'difference_stationary' (ADF rejects AND KPSS also rejects — conflicting; difference the series). Plus two explicit non-ambiguous cases: 'constant' for a zero-variance series, and SQL NULL (never a misleading label) for a degenerate or too-short series. regression := 'ct' applies to both component tests, so a deterministic-trend series that reads non_stationary under 'c' typically reads trend_stationary under 'ct'.


Residual-adequacy set

Apply to residuals from a fitted model (forecast errors), not to the raw series.

ts_ljung_box / ts_ljung_box_by — autocorrelation test (white-noise check, model-aware)

H0: residuals are white noise (no autocorrelation). Reject (p < 0.05) → residuals are correlated; model under-fits.

-- Grouped macro (sourced from residuals.sql)
SELECT series_id, (lb).statistic AS q_stat, (lb).p_value, (lb).lags
FROM ts_ljung_box_by('resids', series_id, ds, e) AS t(series_id, lb)
ORDER BY series_id;

Pass fitted_params := p + q for ARMA(p, q) residuals (or any count of estimated model parameters): df = lags - fitted_params, and the p-value matches R Box.test(type = "Ljung-Box", fitdf = p + q) exactly (default fitted_params := 0 keeps df == lags, correct only for raw/unmodeled residuals).

SELECT series_id, (lb).df, (lb).p_value
FROM ts_ljung_box_by('resids', series_id, ds, e, lags := 10, fitted_params := 1) AS t(series_id, lb)
ORDER BY series_id;

fitted_params at or above the effective (resolved/clamped) lag count is a SQL error naming fitted_params and "degrees of freedom" — it never silently floors to df = 1.

ts_durbin_watson / ts_durbin_watson_by — first-order autocorrelation

DW statistic near 2 = no autocorrelation; < 2 = positive; > 2 = negative.

-- Grouped macro (sourced from residuals.sql)
SELECT series_id, (dw).statistic, (dw).interpretation
FROM ts_durbin_watson_by('resids', series_id, ds, e) AS t(series_id, dw)
ORDER BY series_id;

ts_jarque_bera / ts_jarque_bera_by — residual normality

H0: residuals are normally distributed. Reject (p < 0.05) → non-normal (may affect interval coverage).

-- Grouped macro (sourced from residuals.sql)
SELECT series_id, (jb).statistic, (jb).p_value, (jb).skewness, (jb).excess_kurtosis
FROM ts_jarque_bera_by('resids', series_id, ds, e) AS t(series_id, jb)
ORDER BY series_id;

ts_residual_diagnostics / ts_residual_diagnostics_by — combined adequacy report (recommended entry point)

Runs Ljung-Box, Durbin-Watson, and Jarque-Bera internally; returns a single adequate BOOLEAN verdict plus component fields.

-- Grouped macro (sourced from residuals.sql)
SELECT series_id, (rd).lb_p_value, (rd).dw_interpretation, (rd).adequate
FROM ts_residual_diagnostics_by('resids', series_id, ds, e) AS t(series_id, rd)
ORDER BY series_id;

Diagnostics gotchas

  • Stationarity tests need sufficient length: ADF / KPSS require at least ~20-30 observations for reliable p-values. On very short series, results are inconclusive by default.
  • Apply residual tests to residuals, not raw values: Pass (y - yhat) as the value column, not the original series — ts_ljung_box on raw series will almost always reject, which is expected and uninformative.
  • ts_adf_by output column names are the group column name (preserved) plus the adf STRUCT. Access sub-fields with (adf).statistic, etc.
  • KPSS H0 direction is opposite to ADF: KPSS non-rejection means stationary; ADF non-rejection means non-stationary. Use ts_stationarity_by to avoid this confusion.

Feature extraction (117 tsfresh-inspired features)

ts_features_by — extract all 117 features

ts_features_by(source VARCHAR, group_col COLUMN, date_col COLUMN, value_col COLUMN) → TABLE

Returns the group column + 117 feature columns: mean, standard_deviation, skewness, kurtosis, length, linear_trend_slope, autocorrelation_lag1, …

Every feature matches tsfresh 0.21.2 (or a direct numpy/scipy reference for the 9 features with no tsfresh counterpart) on a 9-series fixture suite, or carries a documented, intentional deviation — see docs/api/21-feature-reference.md's Deviation column.

-- Extract all features
SELECT * FROM ts_features_by('sales', product_id, ds, y);

-- Filter by feature values
SELECT product_id, mean, linear_trend_slope
FROM ts_features_by('sales', product_id, ds, y)
WHERE length > 30 AND autocorrelation_lag1 > 0.5;

ts_features (scalar)

Scalar over (date, value). Same 117 outputs as struct fields.

ts_features_list / ts_features_table

Discover the feature catalogue:

SELECT column_name, feature_name, parameter_suffix
FROM ts_features_list()
WHERE feature_name LIKE '%autocorr%';

ts_features_config_template — starter config

Table function (no args) returning a template config row for feature extraction — a scaffold to edit down into a custom feature subset (feeds ts_features_config_from_json / ts_features_config_from_csv below).

SELECT * FROM ts_features_config_template();

Custom feature subsets

ts_features_by/ts_features_table take no config argument (4/3 positional args only). Configure via JSON or CSV against the ts_features_agg aggregate's feature_selection/feature_params arguments instead, unpacking the config scalar's feature_names/overrides fields:

-- Load a JSON config
SELECT product_id,
    ts_features_agg(ds, y,
        (ts_features_config_from_json('{"features": ["mean", "std", "autocorrelation_lag1"]}')).feature_names,
        (ts_features_config_from_json('{"features": ["mean", "std", "autocorrelation_lag1"]}')).overrides
    ) AS feats
FROM sales GROUP BY product_id;

-- Or CSV
SELECT product_id,
    ts_features_agg(ds, y,
        (ts_features_config_from_csv('mean,std,skewness')).feature_names,
        (ts_features_config_from_csv('mean,std,skewness')).overrides
    ) AS feats
FROM sales GROUP BY product_id;

Note: in this build, the "features" allow-list in the JSON/CSV config did not actually narrow the returned feature set (every call above still returned all 117 features) — logged as a deferred finding, not fixed by this docs sweep.

ts_features_agg — aggregate variant

For custom GROUP BY:

SELECT product_id, ts_features_agg(ds, y).mean, ts_features_agg(ds, y).autocorrelation_lag1
FROM sales GROUP BY product_id;

Gotchas

  • ts_stats and ts_data_quality are TABLE functions — call from FROM, not SELECT. Older docs incorrectly showed them as scalars over LIST(...).
  • _agg variants take (date, value) directly — no LIST() wrapping. Wrapping in LIST(...) errors as "No function matches …".
  • trend_strength / seasonality_strength formulas (v2026.10.16, STAT-02, D-02): both are Wang/Hyndman STL decomposition strengths from an R-faithful STL — trend_strength = max(0, 1 - Var(R)/Var(T+R)), seasonality_strength = max(0, 1 - Var(R)/Var(S+R)) — computed via stats::stl (one period), the forecast::mstl iteration (several periods, reporting the maximum per-period strength), or a supsmu trend when no period is usable. The period comes from the new seasonal_period parameter (0 auto-detects, 1 forces non-seasonal, >= 2 is explicit). seasonality_strength is NULL when no period is usable (none detected, seasonal_period := 1, or n <= 2 * seasonal_period) or when the series decomposes to zero seasonal variance (e.g. a constant series). Rule-of-thumb thresholds (heuristics, not statistical tests — validate against your own data): below 0.3 weak, 0.3 to 0.6 moderate, 0.6 and above strong. Example: the audit's 150-point random walk gives seasonality_strength about 0.09 (weak, matching "mostly noise"); AirPassengers gives about 0.94 (strong, matching its well-known yearly cycle).
  • JSON extension is required by a subset of quality / feature functions. Enable auto-load once per session: SET autoinstall_known_extensions=1; SET autoload_known_extensions=1;.
  • ts_r2(DOUBLE[] actual, DOUBLE[] predicted) → DOUBLE is a scalar regression R² metric, not an EDA statistic — it lives with the scalar metrics family in anofox-forecast-backtest. Use it under GROUP BY to score fit quality, e.g. SELECT ts_r2([1.0,2,3,4], [1.1,2.0,2.9,4.2]);.
  • ts_features_agg / ts_features / ts_features_by share one Rust NULL policy (Phase 12, D-11): a NULL-valued row is no longer spliced out independently — leading/trailing NULLs are trimmed, an interior NULL is a hard error naming its 1-based position, and an all-NULL series returns NULL (the whole struct for the aggregate, a row of NULL feature columns for the table function). All three surfaces agree on the same series.

Canonical EDA pipeline

-- 1. Quality gate
CREATE OR REPLACE TABLE quality AS
SELECT * FROM ts_data_quality('raw', product_id, ds, y, 14, '1d');

-- 2. Series-level stats
CREATE OR REPLACE TABLE stats AS
SELECT * FROM ts_stats('raw', product_id, ds, y, '1d');

-- 3. Filter — keep only high-quality, forecastable series
CREATE OR REPLACE TABLE forecastable AS
SELECT r.*
FROM raw r
JOIN quality q ON r.product_id = q.unique_id  -- ts_data_quality renames the group col to unique_id
JOIN stats  s ON r.product_id = s.product_id
WHERE q.overall_score >= 0.6
  AND s.length >= 30
  AND NOT s.is_constant;

-- 4. Extract features for ML (optional)
CREATE OR REPLACE TABLE features AS
SELECT * FROM ts_features_by('forecastable', product_id, ds, y);

See also: anofox-forecast-data-prep (impute + drop bad series based on these stats), anofox-forecast-detection (trend_strength / seasonality_strength inform detection thresholds), anofox-forecast-models (use quality score / features to gate model selection).

Reference docs:

  • docs/api/03-statistics.md
  • docs/api/20-feature-extraction.md
  • docs/api/21-feature-reference.md