Interpreting Confidence Bands

What the Bands Mean

The forecast chart displays two sets of confidence bands around the point prediction:

• 50% confidence interval (inner, darker band): The actual price should fall within this range approximately half the time. Think of it as the “most likely” price range.
• 90% confidence interval (outer, lighter band): The actual price should fall within this range approximately 9 out of 10 times. Think of it as the “almost certainly within” range.

                    ┌── 90% CI upper (5% chance above)
                   ╱
        ┌── 50% CI upper (25% chance above)
       ╱
──────●──── Point prediction (best estimate)
       ╲
        └── 50% CI lower (25% chance below)
                   ╲
                    └── 90% CI lower (5% chance below)

Reading the Chart

Narrow Bands = High Confidence

When bands are tight around the point prediction, the model is confident. This typically occurs during:

• Night hours (00:00–06:00) when demand is stable
• Day-ahead forecasts (shorter horizon)
• Periods of stable weather and moderate demand

Wide Bands = High Uncertainty

When bands spread far from the point prediction, uncertainty is elevated. This typically occurs during:

• Peak hours (10:00–14:00, 18:00–21:00)
• Longer horizons (D+5 to D+7)
• Extreme weather events
• Market transition periods (holiday edges, season changes)

Asymmetric Bands = Directional Risk

When the upper band extends much further than the lower band, there’s more upside risk (potential for prices to spike above the forecast). This is common during:

• Evening demand peaks (solar dropout + demand rise)
• Low wind periods (risk of expensive thermal dispatch)
• System stress events (maintenance outages, tight margins)

Practical Applications

For Traders

Use the confidence intervals to assess trading risk:

• Buy/sell decisions: If the current market price is below the 90% CI lower bound, the model considers this price extremely unlikely — it may represent a buying opportunity
• Position sizing: Wider intervals suggest larger potential price swings, warranting smaller position sizes or tighter stop-losses
• Scenario planning: The 50% CI bounds define the “expected” range for operational planning

For Grid Operators

Use intervals for operational planning:

• Reserve scheduling: The 90% CI upper bound indicates the worst-case price scenario for dispatching reserves
• Demand response: Wide upward bands signal potential price spikes where demand response activation may be needed

For Renewable Developers

Use intervals for revenue estimation:

• Revenue projections: Use the 50% CI for expected revenue, 90% CI for stress testing
• Curtailment risk: Wide downward bands (low lower bound) suggest risk of negative prices and potential curtailment
• PPA pricing: Long-term confidence bands inform power purchase agreement negotiations

For Risk Managers

Use intervals for Value-at-Risk calculations:

• 90% CI lower bound ≈ VaR(95%): Only 5% chance the actual price falls below this level
• Portfolio exposure: Sum of (position × CI width) across hours gives total price exposure

Common Pitfalls

Pitfall 1: Treating the Point Prediction as Certain

The point prediction is the single most likely outcome, but prices frequently land outside it. Always consider the bands:

Point prediction: 55 EUR/MWh
50% CI: [48, 62]
→ There's a 50% chance the actual price is NOT between 48 and 62

Pitfall 2: Ignoring Asymmetry

If the 90% CI is [40, 85] around a prediction of 55, the risk is not symmetric:

Downside risk: 55 - 40 = 15 EUR/MWh
Upside risk:   85 - 55 = 30 EUR/MWh
→ Upside surprises are twice as likely to be extreme

Pitfall 3: Expecting Exact Coverage

The 90% CI should contain ~90% of actual prices on average, over many predictions. On any single day, it either contains the actual or it doesn’t. Coverage rates are statistical properties, not per-prediction guarantees.

Pitfall 4: Comparing Intervals Across Horizons

A narrow 90% CI at D+1 and a wide 90% CI at D+7 both provide 90% coverage — the D+1 interval isn’t “better.” It’s narrower because D+1 is inherently more predictable. Compare intervals within the same horizon, not across horizons.

Coverage Expectations

Over a 30-day evaluation period with 24 hourly predictions per day (720 total):

Interval	Expected Outside	Actual Range
50% CI	~360 prices outside (50%)	320–400 (normal)
90% CI	~72 prices outside (10%)	36–108 (normal)

If significantly more actual prices fall outside the 90% CI (e.g., 150 out of 720), the intervals may need recalibration.