Forty projects are not
forty coin flips.

A chance for one project answers one question. Will this one get built? It does not answer the question that matters. How much will my forty deliver? To get from the first to the second, almost everyone multiplies. Multiplying assumes each project stands or falls on its own.

Projects in the same queue share things. A grid operator’s study cycle. A transmission owner’s capital plan. An equipment supplier. A state’s policy. When one of those turns bad, every project sharing it moves at once. That is a shared shock: one event, many projects. Multiplying cannot represent it, so it hands you a range far too narrow.

A model of shared shocks will invent them where there are none. So I tested first. Do build counts inside a group swing more than they would if every project stood alone?

Count how many projects got built in each real group. Then shuffle which project sits in which group, keeping the same chances and group sizes, and count again. Do that 4,000 times.

The shuffling matters. The underlying model is off: when it says 65% it means 49%. A textbook test would read that error as projects failing together, and find an effect that was not there. The shuffled version carries the same error, so only the real effect survives.

The middle two columns measure how far build counts swing from group to group. Bigger means more swing. The last divides real by shuffled.

A shuffled swing of 1.0 would mean the model’s chances were exactly right. The shuffled swings run 1.19 to 8.77 instead. That gap is the model’s own error. It is also the size of the mistake a textbook test would have made, since a textbook test expects 1.0.

No shuffle ever reached the real swing, on any row. Chance alone would do that less than one time in 4,000 — written 0.00025, and the smallest number 4,000 shuffles can report. It makes luck a poor explanation. It says nothing about the size of the effect.

Tariffs, cell prices and turbine lead times hit every project of a type at once, wherever it sits. Forty projects across ten states but all solar are far less spread out than they look.

That they fail together is real. It does not follow that my model of it predicts better. So I picked the model by testing, not by chasing the biggest effect.

Draw sets of 40 from years the model never saw. Ask each model for a range it calls 80% sure. Count how often the truth lands inside. The target is 80%. A model that scores below 80% is overpromising. A model that scores above 80% is vaguer than it needs to be.

The usual arithmetic is tolerable on a mixed set and dangerous on a concentrated one. Concentrated is the real case for a developer or lender betting on one technology. The range is meant to miss one time in five. It misses nearly two times in five.

Shared shocks fix that row. They overshoot the other two: 89.0% and 88.1% against a target of 80%. Those ranges are wider than they need to be. Too wide is safer than too narrow. It is still not right.

All 1,636 requests still waiting across four grid operators, 355 GW of capacity. Both models expect about 102 GW delivered. They disagree completely about how sure anyone should be.

Same centre. 4.3× the spread. The bad case is 21.4 GW worse than the usual arithmetic implies. That is about a fifth of the 102 GW expected. It is the number a lender sizing a loan, or a manufacturer sizing a factory, gets wrong.

Two problems, found while building this. Neither is fixed.

Getting the project count right does not get the gigawatts right. Tune the chances until the project count comes out right. Do that, and delivered capacity comes out 20.8% too high on projects the model never saw. The gigawatts sit in the big projects, and big projects fail more often.

Re-tuning by megawatts instead cuts that to 6.1%. Telling the model how big each project is does not help. Compare only projects of a similar size and the bias is still there.

The live queue is 68% MISO by capacity. The history I trained on is 19% MISO. So the headline leans hard on one operator’s record being typical. Split the history by operator and technology, and some slices hold fewer than 40 finished requests. Those thin slices cover 18% of the queue’s capacity.

The 4.3-times-wider figure divides one spread by another, and it survives both problems. The centre does not. So the claim is deliberately narrow. Not “103 GW will be delivered”. Rather: whatever the centre turns out to be, the spread around it is more than four times what the usual arithmetic tells you.

Every project keeps its own chance of being built. On top of that, each group gets one shared push, up or down, applied to all its members. Small pushes are assumed to be commoner than large ones, in the usual bell shape. I fit the settings that make the real outcomes most likely, allowing for every push a group might have had.

The per-project chances are frozen at my earlier model. That model was tested by training on old years and predicting later ones. Freezing it stops the shared push from soaking up ordinary error.

Study cycle measures 1.06 on the scale the model works in, where 0 would mean no shared push at all. Put another way: 25.5% of the variation the per-project model leaves unexplained sits between study cycles rather than within them.

Groups were fitted one at a time. A project belongs to several at once, so their effects overlap and cannot be added up. So I chose the combination by testing on unseen years.