KEYNUM.AI

Energy & infrastructure intelligence

Loading Direction 1 control room

Preparing dashboard snapshot...

Direction 1 · Rule-based evidence layerFrance · EEX rolling front-month

Keynum.ai energy & infrastructure intelligence

Can anomalous behavior signal trouble ahead?

Some signals carry useful warning information, but strength is uneven. Spread structure leads at the main two-week horizon; narrative pressure is most useful over one week.

Historical, out-of-sample classification evidence. This is a transparent warning system, not a deterministic crisis forecast.

Evidence window2024-02 → 2026-05Expanded historical panel

Daily observations839Calendar decision rows

Walk-forward folds22Purged, held-out evaluation

Rule families6One transparent stress channel each

h14 event rate9.8%Background rate in held-out rows

Liquidity Deterioration

Parkinson volatility-over-volume spread proxy

Secondary

Question under test

Can unusually fragile front-contract trading conditions precede a future stress-driven upward repricing episode?

Why this signal

A market with high range-based volatility relative to traded volume may be less resilient to the next shock, even before outright prices move dramatically.

Current reading

A modest fragility signal rather than a decisive warning rule. Results are positive but sparse and unstable across real-bar and all-bar variants.

Anomaly constructionScore and decision rule

Train-fold q90 · leakage-safe

\sigma_t^{P}=\sqrt{\frac{8}{\pi}}\log\!\left(\frac{H_t}{L_t}\right),\qquad S_t^{\mathrm{liq}}=\left(\frac{\left(\sigma_t^{P}\right)^2}{V_t^{\mathrm{front}}}\right)^{1/3}

View LaTeX

$$\sigma_t^{P}=\sqrt{\frac{8}{\pi}}\log\!\left(\frac{H_t}{L_t}\right),\qquad S_t^{\mathrm{liq}}=\left(\frac{\left(\sigma_t^{P}\right)^2}{V_t^{\mathrm{front}}}\right)^{1/3}$$

X_t^{\mathrm{liq}}=\mathbf{1}\!\left\{S_t^{\mathrm{liq}}\ge q_{0.90}^{\mathrm{train}}\!\left(S^{\mathrm{liq}}\right)\right\}

View LaTeX

$$X_t^{\mathrm{liq}}=\mathbf{1}\!\left\{S_t^{\mathrm{liq}}\ge q_{0.90}^{\mathrm{train}}\!\left(S^{\mathrm{liq}}\right)\right\}$$

h77 calendar days h1414 calendar days h3030 calendar days

AUC-PR

0.172

Baseline 0.140

Lift over base

1.23×

Ranking value vs background

Precision

26.1%

6 true-positive dates

Recall

10.0%

54 future events missed

F1 score

0.145

Precision / recall balance

Signal days

Variant: vov spread real

Inspect classification detail

Observed future outcome

Tail event

No tail event

Rule fired

6True positive

17False positive

Rule quiet

54Missed event

353True negative

Held-out observations shown: 430. “Positive” means the anomaly rule fired before the selected future horizon; it does not imply causation.

Analyst caveat

The preferred real-bar proxy is available on 574 of 839 panel rows, so this family has a thinner effective sample than the others.

How to read lift

A lift of 1.23× means the continuous anomaly score ranks future tail-event dates 1.23 times better than the event base rate. It is evidence of association, not proof that the anomaly caused the event.

Rule-based evidence matrix

Canonical real-control result selected for each family and horizon. Full detail stays available without crowding the decision surface.

Family	Horizon	Variant	AUC-PR	Base	Lift	Precision	Recall	F1
Price Spike	h7	abs log	0.122	0.083	1.47×	17.6%	16.7%	0.171
Price Spike	h14	abs log	0.134	0.098	1.37×	15.7%	12.7%	0.140
Price Spike	h30	abs log	0.170	0.148	1.15×	20.8%	10.8%	0.142
Volatility Jump	h7	raw rv	0.125	0.083	1.51×	16.1%	16.7%	0.164
Volatility Jump	h14	raw rv	0.125	0.098	1.28×	14.5%	12.7%	0.136
Volatility Jump	h30	raw rv	0.165	0.148	1.12×	25.0%	14.0%	0.179
Spread / Contract Alignment	h7	cross country max	0.166	0.083	2.01×	13.1%	37.0%	0.193
Spread / Contract Alignment	h14	m1 m3 abs	0.146	0.098	1.49×	20.8%	39.7%	0.273
Spread / Contract Alignment	h30	cross country max	0.189	0.148	1.28×	24.6%	34.4%	0.287
Liquidity Deterioration	h7	vov spread real	0.138	0.085	1.61×	16.0%	10.5%	0.127
Liquidity Deterioration	h14	vov spread all	0.130	0.098	1.33×	20.8%	11.6%	0.149
Liquidity Deterioration	h30	vov spread real	0.172	0.140	1.23×	26.1%	10.0%	0.145
Market-Implied Coupling Pulse	h7	coupling scaled max	0.116	0.083	1.40×	12.0%	11.1%	0.115
Market-Implied Coupling Pulse	h14	coupling scaled max	0.136	0.098	1.40×	18.0%	14.3%	0.159
Market-Implied Coupling Pulse	h30	it spot futures basis snap	0.194	0.148	1.31×	24.1%	14.0%	0.177
GDELT Narrative Pressure	h7	cwt	0.203	0.082	2.48×	18.2%	26.9%	0.217
GDELT Narrative Pressure	h14	log cwt	0.114	0.100	1.13×	8.6%	14.3%	0.107
GDELT Narrative Pressure	h30	log cwt	0.137	0.152	0.90×	6.9%	7.5%	0.072

How classification works

One transparent test, repeated fold by fold.

01
Observe today
Build one continuous anomaly score from data available by the decision cutoff.
02
Learn “extreme”
Estimate the anomaly cutoff from the training fold only.
03
Look forward
Measure cumulative rolling-front return over the next H days.
04
Test out of sample
Compare the fired rule with the held-out future tail-event label.

Downstream trouble target

Empirical upper-tail VaR event, α = 0.10.

R_{t,H}^{\mathrm{cum}}=\prod_{i=1}^{H}\left(1+r_{t+i}^{\mathrm{roll}}\right)-1

View LaTeX

$$R_{t,H}^{\mathrm{cum}}=\prod_{i=1}^{H}\left(1+r_{t+i}^{\mathrm{roll}}\right)-1$$

Y_{t,H}^{\mathrm{up}}=\mathbf{1}\!\left\{R_{t,H}^{\mathrm{cum}}\ge \widehat{VaR}_{H,0.10}^{\mathrm{up,train}}\right\}

View LaTeX

$$Y_{t,H}^{\mathrm{up}}=\mathbf{1}\!\left\{R_{t,H}^{\mathrm{cum}}\ge \widehat{VaR}_{H,0.10}^{\mathrm{up,train}}\right\}$$

The target uses only returns from t+1 through t+H. Boundary-crossing training rows are purged before thresholds are estimated.

Evidence posture

What this demo can and cannot claim.

Can say: some anomaly families rank later tail-event dates above the background rate.
Cannot say: a fired anomaly caused a geopolitical or energy-security event.
Next layer: combine frozen family scores in a purge-safe logistic model.