Crews Idle in February, Work Turned Away in May: 2026 Models Finally Read Storm-Lumpy Roofing Demand

Eleven weeks. That is how long the window can stay open after a serious hail swath crosses a metro area — roughly from the storm to the point where the last insured homeowner has picked a contractor. A roofing company in the Denver, Dallas or Oklahoma City market can book more than half a year of revenue inside those eleven weeks, and then spend October through February trying to keep three crews fed on repairs, gutters and the occasional retail replacement.

That pattern is why roofing demand forecasting has always been treated as a joke in this trade. You cannot forecast hail. Fair enough. But the thing owners actually need to decide is not "will it hail." It is: how many crews do I commit to for the next ten weeks, what do I pay them per square, how much material do I pre-buy before the March price increase, and do I hold or discount my retail price this month. Those decisions get made on gut feel, every year, and getting them wrong costs more than any single job on the board.

The four guesses that set your year

Guess one: crew capacity. Subcontract crews go where the work is. If you commit to three crews and the work comes in for five, you turn away jobs at exactly the moment margin is highest. If you commit to five and the storm goes north of you, you are feeding crews you cannot bill, and they leave anyway.

Guess two: retail price per square. Note the important nuance in this trade — on an insurance-funded replacement you do not really set price, the carrier's price list does, and your lever is the supplement and the accuracy of the scope. Your actual pricing lever lives on the retail side: cash re-roofs, repairs, gutters, siding, and the upgrade lines the homeowner pays out of pocket. Most shops carry one price per square all year, discount in February to keep crews busy, and hold that same number in May when the phone will not stop.

Guess three: the pre-buy. Shingle manufacturers announce price increases with a few weeks of notice, and most shops either buy nothing ahead or buy a warehouse full and tie up cash they need for payroll.

Guess four: where to send the canvassers next week. That one is normally settled by whoever shouts loudest at the Monday meeting.

What changed in 2026 is that demand forecasting models now handle history that is sparse, seasonal and violently uneven — the exact shape of roofing demand — which is why forecasting has become one of the highest-adoption uses of AI in industry, running around 48% in manufacturing, with pricing optimisation higher still at roughly 72% in retail and e-commerce.

Why the old tools choked and these do not

Classic forecasting wants a repeating seasonal curve. Ice cream in July, tax software in April. Roofing does not have that. It has a flat baseline of repairs and retail work, plus enormous spikes that are geographically narrow — one side of a highway gets 1.75-inch hail and the other side gets rain — plus a long tail of claims filed eight and ten months after the event, plus a supplement cycle that pushes revenue recognition weeks past install. Feed that into a spreadsheet trend line and you get nonsense, which is why owners stopped trying.

The 2026 models do not need a clean curve. You can hand them five years of your own job history — every job with its date, zip code, squares, retail or insurance, close rate, and what you charged — alongside the messy outside signals: storm reports by date and location, residential re-roof permits pulled by the jurisdictions you work in, and your own inbound call volume by day. What comes back is not a prophecy. It is a range: how many squares you are likely to install per week for the next ten weeks, by zip code, with a high and low case, and it updates when reality moves.

flowchart TD
  A["Five years of jobs by zip, week and squares"] --> B["Model adds storm reports, permits pulled, call volume"]
  B --> C["Ten-week outlook: squares per week, high and low case"]
  C --> D["Owner sets crew commitments and the pre-buy"]
  D --> E["Actual installs and close rates land back in the job system"]
  E --> A
  C --> F{"Is the install backlog past 21 days?"}
  F -->|"Yes"| G["Hold retail price, stop discounting, add a crew"]
  F -->|"No"| H["Move the canvass team to the next zip"]

What an owner actually does with the number

The output is only useful if it changes a decision on a Monday. Three decisions it changes.

Backlog-triggered pricing. If the outlook says you will be at a 26-day install backlog by the third week of May, the discount your salesman wants to give to close tonight is costing you a better job next week. Shops that put a simple rule on this — backlog over 21 days means no discretionary discounting and a step up on the retail price per square — recover margin without a single uncomfortable conversation, because the rule takes the argument out of it.

Crew commitments in writing. If the low case says 340 squares a week and the high case says 610, you commit to the crews that cover the low case on a standing basis and line up the difference on a callable basis. That is a conversation you have in February for April, not a panic text in May.

Canvass routing. Permits pulled and storm swaths together tell you which zip codes still have unclaimed roofs and which have already been picked over by the out-of-state chasers. That is worth more to the sales manager than any script.

Working the crew-commitment number, with assumptions shown

Illustrative shop: Dallas–Fort Worth, residential, mixed insurance and retail, average job 31 squares, average contract value $16,300, gross margin 30%. A crew installs about 5 roofs a week in season.

Cost of under-committing: 60 jobs × $16,300 × 30% = $293,400 of gross profit that walks to the contractor whose ladder was already up. Cost of over-committing is subtler — you rarely pay an idle subcontract crew directly, but you lose them, and re-recruiting a good crew in the middle of a season costs you weeks of production and a higher per-square rate. Forecast-guided lands about 190 of the 210, worth roughly $929,000 of gross profit against $733,000 in the under-committed case.

That gap is not a claim that a model conjures $200,000. It is a claim that the difference between three crews and five, decided six weeks earlier with a range rather than a hunch, is worth that much. Run the same table on your own average job value and crew throughput before you believe any of it.

Where the forecast will lie to you

It cannot see the storm that has not happened. Nothing can. Anyone selling a roofing demand model on the promise of predicting hail is selling you a weather forecast with extra steps. What it predicts is the conversion of demand you can already observe — claims filed, permits pulled, calls received, doors knocked — into squares installed on your boards.

It is weak in a market you just entered. Five years of your own history in Tulsa tells the model nothing useful about Wichita. Expect the first season in a new metro to be run on judgment, with the model watching.

And it will not tell you whether your crews can actually hold quality at the high case. That is the failure that ends roofing companies — taking 210 jobs with crews you have not worked with, generating a callback rate that eats the season's profit in year two. The forecast is a demand number, not a capability number. Your production manager owns that one, and he should have veto power over the growth case regardless of what any model says.

Frequently asked questions

We only have three years of clean job history. Is that enough?

Usually yes, for a range rather than a point. Three years covering at least one significant storm event and two normal winters gives a model enough to separate your baseline repair and retail work from event-driven spikes. What matters more than the number of years is whether the zip code and the square count were recorded on every job, which in many shops is the actual gap.

Does this let us raise prices without losing bids?

On the retail side, in a peak backlog, often yes — a homeowner who has already called four contractors and been told eight weeks is not shopping the last $400. In February, no. The value is knowing which month you are in and stopping your sales team from discounting in the wrong one.

Our jobs are 80% insurance. Is there anything here for us?

The pricing half matters less, because the carrier's price list sets it. The capacity half matters more, because insurance work arrives in the sharpest spikes of anything in the trade. Forecast squares and crews; leave price alone and put that energy into supplement accuracy instead.

What does this cost to run?

Far less than owners expect. Running a weekly forecast off your own job history and public storm and permit data is a small recurring cost now that capable models sit around a couple of dollars per million words of text processed — the real expense is the two or three days of somebody cleaning up five years of job records so the zip codes and square counts are actually there.

One practical note on the spike itself: the week after a storm, inbound calls can run five to ten times normal, and the shops that win are simply the ones that answered. CallSphere builds AI voice and chat agents that pick up every one of those calls at any hour, take the address and damage details, and book the inspection straight onto an estimator's calendar — which also gives you a clean daily call count by zip code, the single best early signal your forecast can have.