Offline tests
At the publication checkpoint, including the recorder recovery. Network access is denied during the Python test phase. The synthetic demo is explicitly invented data.
ENERGY FORECAST LAB RESEARCH NOTE / 2026.09
Forecast the peak.
Price the
uncertainty.
A better forecast is only the beginning of a trading strategy.
I built a research system for Texas electricity demand: correct the operator’s errors, model the range of possible peaks, then test whether an executable price can turn that information into value.
Forecast improvement measured.
Trading profitability unproven.
01 / THE FULL YEAR
Every day stays in the picture.
Explore daily peak forecasts from the intraday experiment. Change the clock, model or month; the missing dates remain visible.
Revised observed peakForecast median90% model interval
F002: 361 scored days out of 365 calendar dates.
View the selected observations as a table
| Date | Actual MW | Median MW | 90% range MW | Label |
|---|
F002 · These are forecasts of a revised EIA demand proxy, with assumed historical arrival clocks. They are not first-release ERCOT settlement forecasts or a price backtest. The displayed intervals are F002’s original distributions; the calibration experiment below uses separate, matched controls.
12.37%
Lower daily peak error
F001 ridge vs operator. 357 common days; 4.04% improvement over the stronger mean-correction baseline.
19.48%
Lower afternoon error
F002 ridge vs half update at 15:00. 361 days; descriptive comparison on an inspected development year.
15.04%
Better interval score
F003 scaled ridge vs matched ridge at 14:00. A score that penalizes both wide ranges and missed outcomes.
02 / THE RESEARCH QUESTION
Predict the error
in a strong forecast.
Electricity demand has a useful starting point: the grid operator already publishes a forecast.
The first task was to estimate its remaining error. A model that learns persistent bias and incorporates the day’s observed trajectory can produce a distribution for the final daily peak.
The trading target is more specific: will the contract’s official hourly peak strictly exceed a threshold? The settlement source is ERCOT’s Actual System Load by Forecast Zone, NP6-346-CD. Revised EIA history is useful for model development, but it does not establish what that original release said.
This project grew out of WeatherPred, where better weather forecasts also failed to establish a trading edge. The energy work carries forward the timestamp, execution and validation lessons. Its implemented models are regression and residual scenarios; transformer and reinforcement-learning ideas are not energy results.
01
Receive
Archive inputs and actual receipt clocks.
02
Correct
Estimate the operator’s remaining error.
03
Simulate
Keep whole hourly residual paths together.
04
Price
Compare probabilities with executable asks.
05
Reconcile
Account for fills, fees, cash and settlement.
03 / WHAT THE MODELS ACTUALLY LEARNED
Simple corrections
do most of the work.
About 70% of the initial gain came from correcting average bias.
The raw operator peak had a +905 MW average bias in the common evaluation sample. Mean correction reduced that to +317 MW; ridge reduced it to +149 MW.
Ridge lowered mean absolute error from 1,627 to 1,426 MW. The relevant second comparison is 1,486 MW for mean correction. Complexity adds value here, but much less than the headline comparison alone suggests.
Timing beats
a blanket update rule.
At noon, ridge loses to the calibrated operator. By 15:00, it is the strongest of these four methods on all three reported average scores. The right lesson is to test the information available at each decision time.
MAE measures the average absolute error of the predicted peak; lower is better.
Exact results for all clocks and models
| Clock | Model | MAE MW | CRPS MW | Brier |
|---|
F002 · 361 distinct scored days per clock, not thousands of independent outcomes. The 15:00 ridge-vs-half headline is descriptive and was not a registered inferential contrast. Download the complete three-metric figure.
04 / ACCURACY IS NOT CALIBRATION
A precise-looking forecast
can be too confident.
A nominal 90% range covered only 81.72% of outcomes at 14:00 in the matched ridge control.
Scaling the residual paths raised coverage to 90.30%. The ranges became 23.07% wider, but a score charging for both width and misses improved by 15.04%.
The improvement was narrower than “the model is calibrated.” Its full-distribution score improvement had an adjusted interval crossing zero. At 15:00, even the scaled nominal 95% interval covered only 91.97% of outcomes.
The selection step is part of the experiment.
Each month and clock uses separate fitting, donor-library and tuning windows. A scale of 1, 1.25, 1.5 or 2 is chosen before that month’s evaluation. Scale 1—no widening—wins in many folds.
All 24 comparisons are retained. Shared seven-day blocks account for serial dependence when summarizing uncertainty; adjusting the intervals does not make a reused development year an untouched holdout.
WHOLE-PATH DISPERSION
e′j = ē + s(ej − ē)
Spread each historical hourly error path around the mean path, then take the daily maximum. Hourly means stay fixed; the nonlinear peak’s median can move.
Read the derivation →Inspect all monthly scale choices
Inspect all 24 comparison intervals
05 / THE FORECAST HAS TO SURVIVE EXECUTION
There is still a gap
between signal and profit.
The first price screen found no positive guaranteed package surplus.
For two nested thresholds, YES on the lower threshold plus NO on the higher one pays at least $1 in every outcome. A buy price below $1 would be a necessary first condition for that particular package to lock in gross value.
Across 1,980 adjacent-pair observations, only 127 had both purchasable legs. None passed that condition before costs. These were repeated snapshots of one daily outcome, not 1,980 independent opportunities. This result rejects the observed packages, not every directional strategy.
FORWARD PAPER PILOT / FROZEN POLICIES
18 alternative $200 accounts.
Three forecast rules × two risk budgets × three execution assumptions. These are parallel policy comparisons, not $3,600 invested.
- Entry
- Stressed probability edge after ask, fees and a price buffer.
- Size
- Whole-contract expected-log sizing, capped by cash, risk and displayed depth.
- Fill
- Later order books, partial fills, slippage and conservative depth reductions.
- Cash
- Unresolved attempts keep reserves locked; settlement uses finalized outcomes.
Publication checkpoint · 7 September 2026, before the first decision
Zero completed decisions, orders or settlements. No return, win-rate or drawdown claim is available. The first decision is scheduled for 12:00 CT. An overnight quote-recorder failure caused a real coverage gap; recovery is documented separately and does not backfill missed observations.
06 / ENGINEERING THE EVIDENCE
A result should be
reconstructable.
The system records raw response receipts, distinguishes event time from arrival time, freezes experiment definitions and reconstructs outputs independently. Missing inputs and unsuccessful attempts remain visible.
The public repository provides the energy modules, tests, synthetic execution example, retained result reports and chart data. It reproduces the figures and checks the software offline. Full historical experiments additionally require the separately retained raw inputs; the included reports alone cannot reproduce those data acquisitions.
Numerical comparisons
The retained F003 independent audit checked predictions, scores, tuning outputs and every bootstrap comparison across 342 artifacts.
Calendar days retained
Missing forecasts, missing observations and invalid retrospective labels are distinguished. Only 361 days per clock are scored in F002/F003.
07 / WHAT CHANGED MY MIND
The baseline deserves serious work.
Most of the initial improvement came from removing bias. Any complex model should clear that stronger baseline before claiming useful incremental skill.
Different decisions need different objectives.
Peak error, probability quality, interval calibration and realized net return answer different questions. An improvement in one does not settle the others.
Access and depth can dominate model choice.
The gas-storage alternative remained blocked by original-release and executable-price data. Wholesale electricity spreads require a separate access route. Neither is a validated result here.
A narrower claim can be more useful.
The current contribution is an auditable forecasting and execution research system. A tradable edge still needs timestamp-correct input and quote overlap, independent evaluation and forward fills.
REPRODUCE / REVIEW / DISCUSSOpen the notebook.
Open the notebook.
Follow the evidence.
Project design, implementation and research direction by Martin Mashalov, developed with AI coding and research assistance. The methods are established statistical tools; the contribution is their integration, measurement and audit trail.