• Wide price dispersion.price_eur_mwh ranges from −22.87 to 308.60 EUR/MWh with stdev 71.32. The mean (107.3) sits well below the median (130.5), indicating a left-skewed distribution dragged down by a cluster of cheap and negative intervals.
• Band distribution is top-heavy.expensive dominates at 676 rows (53%), followed by cheap (295), normal (235), negative (49) and extreme (13). Only ~19% of intervals are "normal" — this is a volatile day by construction.
• Negative pricing is material. 42 rows are flagged is_negative_price (3.3%). Note the discrepancy with the 49 rows in the negative price band — worth reconciling before use, as the two fields apply different thresholds or logic.
• Volatility clustering.is_high_volatility is True for 212 rows (16.7%), while spikes are rarer at 29 rows (2.3%). volatility_6h reaches 111.6 EUR/MWh against a median of 26.04, so extreme regimes are concentrated rather than diffuse.
• Expected rolling-window gaps.volatility_2h has 98 nulls and volatility_6h has 322 nulls (25%) — approximately 23 warm-up rows per zone across 14 zones. price_change_pct has 21 nulls (first observation per zone). One price_change_pct outlier hits +47 (4700%), almost certainly a rebound off a near-zero or negative price; treat percentage changes with care in negative-price regimes.
• Normalisation is clean: price_vs_day_mean and price_zscore both centre on zero, confirming per-day standardisation.

Dataset Context

The dataset (power-price-volatility, energy category) holds 1,268 records of day-ahead spot prices in EUR/MWh across 14 European bidding zones. Zone coverage is balanced at the top of the frequency table — AT, BE, DE-LU, DK1 and DK2 each contribute 96 rows — and the unix_seconds timestamp column spans a narrow window (stdev ≈ 2.49e+04 seconds, roughly a single day of quarter-hourly or hourly slots). Raw prices range from -22.87 to 308.6 EUR/MWh, with mean 107.3 and median 130.5, indicating a left-skewed distribution where a tail of low and negative prices pulls the mean below the median.

Preparation Prior to Derivation

1. Normalize — schema/type harmonisation applied to all columns.
2. Dedupe on the composite key (bidding_zone, unix_seconds), enforcing one price observation per zone per timestamp. This is the key that makes all subsequent per-zone series well-defined.
3. Filter to rows where price_eur_mwh is non-null. Consistent with this, price_eur_mwh is 1268/1268 non-null in the final profile.
4. Sort by bidding_zone, then unix_seconds — a prerequisite for any lag- or window-based statistic, since ordering determines what "previous" and "trailing window" mean.

All derivations below are computed grouped by bidding_zone (by: ['bidding_zone']), so no statistic mixes observations across market areas.

Derived Statistics

Period-over-period return

price_change_pct = pct_change(price_eur_mwh) — the fractional change against the preceding timestamp within the same zone. The first observation of each zone has no predecessor, which accounts for the 21 nulls (1247/1268 non-null). Distribution: mean 0.128, median -0.002851, min -4.996, max 47.0, stdev 1.807. The near-zero median against a positive mean and a maximum of 47 shows the metric is dominated by a few extreme ratios — a structural consequence of dividing by prices that approach zero.

Trailing volatility at two horizons

• volatility_2h = volatility(price_eur_mwh, window=8) — rolling dispersion over 8 preceding observations. 1170/1268 non-null (98 nulls from window warm-up across the 14 zones). Mean 13.27, median 9.289, range 0.07251–60.64 EUR/MWh.
• volatility_6h = volatility(price_eur_mwh, window=24) — the same statistic over 24 observations. 946/1268 non-null (322 nulls, the larger warm-up cost of the wider window). Mean 31.63, median 26.04, range 0.4006–111.6 EUR/MWh.

Both are right-skewed (mean above median). The longer window reports systematically larger dispersion, as expected when a wider span admits more of the intraday price shape.

Rolling level

rolling_avg_2h = rolling_avg(price_eur_mwh, window=8) — trailing 8-observation mean, sharing the volatility_2h window and therefore its 1170/1268 coverage. Mean 104.7, median 128, range -11.89 to 276.9 EUR/MWh. Smoothing compresses the range relative to raw prices (which reach 308.6 and -22.87) and the mean sits close to the raw mean of 107.3.

Centring and standardisation

• price_vs_day_mean = anomaly(price_eur_mwh) — deviation of each price from its zone mean. Fully populated (1268/1268). Mean is -1.22e-15, i.e. zero to floating-point precision, confirming the per-zone centring. Range -134.6 to 196.9 EUR/MWh, stdev 57.12.
• price_zscore = zscore(price_eur_mwh) — the same deviation scaled by the zone standard deviation. Mean -3.856e-17 and stdev 0.9949, both consistent with a correctly standardised variable. Range -2.435 to 3.425; median 0.1664 again reflects the left skew.

Positional index

intraday_position = minmax_normalize(price_eur_mwh) — each price rescaled to [0, 1] against its own zone's observed min and max. Observed bounds are exactly 0 and 1 (each zone contributes its own extremes), mean 0.4615, median 0.5134, stdev 0.312.

Classification and Flags

price_band bins raw price into five ordered, mutually exclusive labels via upper-bound cuts (default unknown, unused in the result):

expensive is the modal band at 676 of 1268 rows (~53%), which is coherent with a median price of 130.5 falling inside that interval.

Three booleans encode threshold events:

• is_negative_price — price_eur_mwh < 0: 42 True, 1226 False.
• is_price_spike — price_zscore > 2: 29 True, 1239 False. Defined on the standardised series, so it is a relative spike within a zone, not an absolute price level.
• is_high_volatility — volatility_2h > 25 EUR/MWh: 212 True, 1056 False.

A validate step closes the pipeline.

Limitations

• Warm-up nulls are unevenly distributed across metrics.volatility_6h is missing for 322 rows (25% of the dataset) and volatility_2h/rolling_avg_2h for 98. Any comparison of the two volatility horizons, or of volatility against price, is drawn from different row subsets unless explicitly restricted to complete cases.
• is_high_volatility inherits those gaps. Rows where volatility_2h is null cannot satisfy > 25, so they fall into the False bucket. The 1056 False count therefore conflates "measured as low volatility" with "not measurable".
• price_change_pct is unstable near zero prices. With prices reaching -22.87 and crossing zero, the denominator can be arbitrarily small; the maximum of 47 (a 4,700% move) and the mean/median divergence are artefacts of this rather than descriptions of typical behaviour. Sign flips across zero also make the percentage uninterpretable in direction.
• Standardisation assumes a meaningful per-zone scale.price_zscore and intraday_position are computed against each zone's own distribution over this single short window (~96 observations per zone). Thresholds like z > 2 are relative to that small sample, so spike counts are not comparable to a longer historical baseline, and intraday_position is anchored to extremes that are themselves sample-specific.
• Short time span limits generalisation. The timestamp stdev of ~2.49e+04 seconds implies roughly one day of coverage. No seasonality, weekday effect, or multi-day mean reversion can be estimated from this data; "day mean" in price_vs_day_mean is effectively the whole-sample zone mean.
• Band cut-points are exogenous. The 0/40/120/250 EUR/MWh boundaries are fixed absolute levels, not data-driven quantiles. They are applied uniformly across all 14 zones despite differing price levels, which is why one band absorbs over half the rows and extreme holds only 13 — small-cell counts unsuited to stratified inference.
• Flag definitions overlap and are not independent.is_negative_price (42) and the negative band (49) are built from the same underlying variable with different comparison operators (< 0 versus ≤ 0), and the seven-row difference reflects that boundary treatment rather than a data discrepancy. Similarly is_price_spike and the extreme band both mark high prices by different constructions and should not be treated as corroborating signals.
• No causal or exogenous content. Nothing in the pipeline includes load, generation mix, interconnector flows or fuel prices, so the metrics describe price behaviour only and support no attribution of its drivers.