Does quantum computing help you trade transmission congestion?
A Financial Transmission Right pays out the price gap between two points on the power grid — the standard hedge against congestion. Three studies, one instrument: forecast what it pays, choose which ones to hold, price an option on one. Each one runs a quantum method against its best classical rival, honestly, on real NYISO data.
Same real data across all three: twelve months of NYISO's public day-ahead price and load reports, all eleven real grid zones, nothing simulated. They're wired together too — the portfolio optimizer actually consumes the forecasting model's output, not a stand-in number. Every result below gets checked against an independent classical reference before I write it down. If the honest answer is a tie, it says tie.
Forecasting the spread
Two quantum kernel methods, tuned through a real hyperparameter sweep, predict tomorrow's day-ahead congestion price between NYISO's NYC and West zones — benchmarked against a tuned classical kernel, gradient boosting, and an LSTM, all on the identical test set with bootstrap confidence intervals on every number.
Result — prediction error, lower is better
In dollars per megawatt-hour. Quantum and classical kernel confidence intervals almost completely overlap. Not a loss, not a win — a real tie.
What actually happened
The plain LSTM won outright, and it's not about quantumness — it's architecture. It sees a full week of raw price history. Every kernel method, quantum or classical, only gets a five-number snapshot of it. That's a real finding on its own: the bottleneck here was how much context the model could see, not which kernel measured similarity between points.
Quantum matches classical. Neither beats a plain LSTM with more context. Architecture was the bottleneck, not kernel choice.
Choosing which contracts to hold
A portfolio optimizer picks 10 of 110 real transmission contracts to hold at once, framed as a constrained optimization problem and solved with a quantum-inspired annealing method, then checked against an exact brute-force solution and a classical benchmark.
What the data itself turned up
At small scale, where I can check the exact answer, both the annealing method and a real gate-based algorithm (QAOA) landed on it dead-on. That's a correctness result — the method works. It's not a speed result: classical annealing solved the same problem faster, with less machinery.
Wiring in a real forecast
The real test: swap one contract's assumed value for the forecasting model's actual prediction, and see if the recommended portfolio moves.
| View of the NYC–West contract | Expected value | Risk |
|---|---|---|
| Full-year historical average | $12.40 | $13.63 |
| Same-period historical average | $15.83 | $19.32 |
| The forecasting model's actual prediction | $8.76 | $33.25 |
The forecast disagreed with history on both counts that matter — lower expected value, higher risk, same window. Feed that in and the optimizer drops the contract entirely, picks something else. Not a relabeling exercise. A forecast that actually changed the answer.
The shape of the problem mattered more than the algorithm did. Using the real forecast changed the answer.
Pricing an option on the spread
Quantum Amplitude Estimation prices a call and a put on the same NYC–West spread, centered on the forecasting model's own prediction rather than a separately-assumed number — and because a congestion spread can go negative (unlike a stock price), the underlying distribution had to be the one real markets already use for exactly that situation, not the standard textbook one.
Result
At the tightest precision tested: over 4 billion classical samples vs. well under a million quantum queries for the same accuracy — but a query is not a sample, and the fourth number above is what happens when you price one.
Why the query gap is bigger than a stock option's
Bigger query gap than the equivalent stock-option result, and the reason tells you something: same quadratic-versus-linear law both times, just more dramatic here because a real congestion spread is noisier than a near-the-money stock option — and classical cost scales straight off that noise.
Finishing the cost math
That 5,000× is a count of queries, and queries aren't free. Each one runs a Grover operator that transpiles to 684 gates on this circuit. Multiplying that through is the same check that killed v2 of the swing-option study, so I ran it here too rather than let this page make the mistake that page is about.
The advantage survives — at 7.4×, not 5,000×. Three orders of magnitude of the headline were per-query cost I hadn't priced in. And at two of the fourteen precisions tested the total drops below 1×: quantum costs more.
The query law holds clean, on real noisy market data — that part is solid. The 5,000× headline doesn't survive its own cost math: 7.4× once queries are priced, and inside the noise of how you count a classical sample.
What it means
Put together, these three studies say something more useful than "quantum works" or "quantum doesn't." They say where it does, on a real problem, checked instead of assumed.
- The advantage isn't uniform, and that's the finding. Pricing showed the query law holding exactly as theory predicts, and a total-cost advantage that nearly vanished once I priced the queries. Forecasting showed a tie. Portfolio selection showed the algorithm working, correctly, but mattering less than a fact about the data itself. All three get reported as measured — not smoothed into one headline.
- Moving from textbook to real broke something every time. Good. A percentage-error metric stopped meaning anything once prices could go negative. A risk-tolerance number tuned for stock returns let risk swamp return by about 100× on raw energy prices. A standard pricing distribution assumed a positivity a real congestion spread doesn't have. I caught every one of these by checking the output against a known-correct answer, not by reasoning it out in advance. That's the actual point of building it this way.
- These pieces are wired together, not just sitting next to each other. The portfolio optimizer actually consumes the forecasting model's output. Small integration, in absolute terms. But it's the difference between three demos next to each other and one system where a prediction can actually change a decision.