Archive / Energy & compute / E829
Past power price spreads do not predict the next day's
In a large US wholesale power market, the recent average gap between day-ahead and real-time hub prices gives no edge for financial (virtual) bids: the holdout earns -$0.01 per MWh after a $1 cost.
US power markets let purely financial traders bet on the gap between the day-ahead price and the real-time price. If utilities overpay to lock in prices a day ahead, that gap should be persistently positive and predictable from its own recent history. On 217,536 hub-hours over three years, the average gap was only $0.34 per MWh, and a rule that follows the recent gap earned -$0.01 per MWh in the holdout year after a $1 cost, and the worst single day lost $22,777.
Day-ahead power prices carry a predictable premium over real-time prices that a financial bidder can collect.
Why people believe it
Utilities and retailers buy most of their power in the day-ahead market to avoid real-time price spikes, and may pay a premium for that certainty. US system operators allow virtual bids: a trader sells day-ahead (an increment) and buys back in real time, or the reverse (a decrement), without owning any generation. If the premium depends on hub, hour and weekday, its recent history should tell a virtual bidder which side to take.
Why it would persist The argument, as stated before testing
Taking the other side means carrying real-time spike risk, needs market membership and collateral, and the premium compensates exactly the risk that hedgers want to shed.
For each hub and hour of the next day, we looked at how day-ahead and real-time prices compared over the previous four weeks (with a one-week lag, so everything was known before the bid deadline). If day-ahead had been more than $1 above real time, we simulated selling day-ahead; if more than $1 below, buying. The verdict was taken on a holdout year, with a control that uses the history of a random other hour.
Holdout mean at least +$0.50 per MWh at c = $1, day-clustered 95% interval above 0, median day above 0, positive at 5 of 8 hubs, increments and decrements each at least 0, control clearly below.
Holdout mean at or below 0 at c = $1, or the result disappears without the 5 best days, or only increments work and the worst day exceeds 30 days of their gains.
c = $1 per MWh primary ($0.50 and $2 reported). Collateral and other operator charges not modelled.
Negative control: the same rule with the signal taken from a random other hour's history.
The holdout earns -$0.01 per MWh at c = $1, with an interval from -$2.70 to +$2.35 and a heavy real-time loss tail.
- Across all 217,536 hub-hours (all 24 hours, 8 hubs, three years), day-ahead exceeded real time by only $0.34 per MWh on average, a third of the assumed $1 cost.
- The rule lost in every window: -$0.61 per MWh in the report year (0 of 8 hubs positive), -$0.31 in validation (1 of 8), -$0.01 in the holdout.
- In the holdout, increments earned +$0.37 and decrements lost $1.03 per MWh; without the 5 best days the result falls to -$1.47.
- The median holdout day lost $36 and the worst lost $22,777 for 1 MW at each of 8 hubs over 13 hours.
- The random-hour control earned +$0.22 per MWh, above the real signal.
Negative or zero in every period
Mean net result per traded MWh after the $1 cost, with day-clustered 95% intervals. The pass bar was +$0.50 with an interval above zero; the holdout interval runs from -$2.70 to +$2.35.
Chart data ($/MWh)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| Report year 2023-24 | -0.61 | -1.23 | 0.02 |
| Validation 2024-25 | -0.31 | -0.93 | 0.34 |
| Holdout 2025-26 | -0.01 | -2.7 | 2.35 |
| Control: random hour's history | 0.22 | Not reported | Not reported |
The holdout, taken apart
Holdout net per MWh after the $1 cost. Only increments are positive, and the total depends on a handful of days.
Chart data ($/MWh)
| Holdout net | |
|---|---|
| All trades | -0.01 |
| Increments (sell day-ahead) | 0.37 |
| Decrements (buy day-ahead) | -1.03 |
| Without the 5 best days | -1.47 |
Why it failed Interpretation · not established
At trading hubs the day-ahead premium has already been competed down to about $0.34 per MWh by professional virtual traders, and its recent history does not say which way the next day will go. What remains is real-time price risk: the median holdout day lost $36, and the worst lost $22,777.
Lesson
A premium can be real on average and still be smaller than the toll for collecting it. Here a near-zero holdout mean (-$0.01) hides a result that rests on a handful of days: without the 5 best days it falls to -$1.47. The day-clustered interval and the without-best-days check expose this; a mean over hours does not.
Scope of this result
Covers 8 hubs of one US system operator, hours 1-13, October 2023 to September 2026, with a trailing-history signal and a $1 cost. It does not cover node-level prices, other operators, hours 14-24 as a verdict, or signals built from information other than past prices.
A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.
1. Test design
Pre-registeredDataUniverseSignalExecutionCostsControlsSample splitPass criterionKill criterionPower2. Result
ObservationsInstrumentsNet mean95% intervalMedianWorstFinding 1Finding 2Finding 3Finding 4Finding 53. Interpretation
Why it failedLessonScope4. Reproducibility
StepsGrade