A GM asks: "How many cars will we ship next month?" You don't know. You guess, based on what's in WIP, what's scheduled, what historical volume looks like, and whatever feels right. You're within 20% most of the time.
The owner asks for a rolling 90-day cash flow. You start with the GM's number, adjust for severity, and hand off a spreadsheet that's half-guesswork.
Forecasting vehicles out—for 30, 60, and 90 days—is one of the highest-leverage models a collision MSO can deploy. Staffing, cash flow, carrier negotiations, and capital planning all depend on it. Here's what a real forecast looks like, what accuracy to expect, and how GMs actually use it.
What Vehicles-Out Forecasting Actually Predicts
The forecast predicts the count of vehicles delivered (out the door, billed) in each future period. Usually:
- 30 days: operational planning (staffing, carrier commitments, bay utilization)
- 60 days: financial planning (cash flow, AR projections)
- 90 days: strategic planning (capital allocation, hiring, M&A pipeline)
These are different forecasts with different sensitivities. A 30-day forecast is mostly about what's already in the funnel. A 90-day forecast is mostly about intake trends.
What a Real Forecast Model Uses
Starting Point: Historical Volume
The first input is simply how many vehicles your shops have delivered in the past. Weekly or daily series, by shop, ideally two or more years of history.
Seasonality
Collision volume has real seasonality. Winter storms. Summer travel accidents. Holiday effects. Back-to-school traffic patterns. Regional weather events.
Good forecasts decompose seasonality: weekly (Mondays vs Fridays), annual (January vs July), and event-based (named storms, holidays).
Current WIP
What's in the shop right now is the strongest predictor of what'll ship in the next 14 days. WIP age distribution, stage distribution (blueprint, parts, repair, paint), and historical completion-by-stage rates convert WIP into near-term delivery predictions.
Intake Pipeline
Scheduled intakes, open estimates, DRP assignments, recent marketing activity. These drive the longer-tail forecast (45–90 days).
External Signals
Miles driven, weather forecasts, local construction activity, carrier relationship changes. These add precision but are optional. A forecast that uses them outperforms one that doesn't by 5–15%.
Model Types That Actually Work
ARIMA / SARIMA
Classical time-series models. Work well on stable historical volume with seasonality. Cheap to run. Documented, explainable.
Best for: shops with at least 18 months of history and limited model complexity.
Facebook Prophet
Designed for business time-series with multiple seasonalities and holiday effects. Handles missing data gracefully. Works well out of the box.
Best for: multi-shop MSOs where you want one model architecture that works across shops with different profiles.
Gradient Boosted Trees (XGBoost, LightGBM)
ML models that can incorporate a wide range of features—weather, pipeline, WIP, external signals. More accurate than ARIMA/Prophet with enough features, less interpretable.
Best for: mature deployments with good feature engineering.
Ensemble
Combining two or three model types into a weighted average. Often beats any single model. Diminishing returns past three components.
Best for: production deployments where you've already maxed out simpler approaches.
For most MSOs at the 10–50 shop range, Prophet + gradient boosted trees in an ensemble covers 90% of the value.
Accuracy You Should Actually Expect
Benchmarks from deployments we've seen:
| Horizon | Mean absolute percentage error |
|---|---|
| 7 days | 5–10% |
| 30 days | 8–15% |
| 60 days | 12–20% |
| 90 days | 15–25% |
These are shop-level forecasts. Aggregated to MSO level, errors tend to be lower (diversification across shops smooths variance).
If a vendor claims single-digit accuracy at 90 days, press hard. Either they're cherry-picking, or they're using a benchmark that doesn't reflect real deployment conditions.
What GMs Actually Do With the Forecast
Staffing
"We'll need 14 tech-days next week based on scheduled work. We have 12 available. We're short."
This is the most common use case. The forecast converts to required tech-hours, compared to scheduled tech availability, and the gap is the GM's staffing problem to solve.
DRP Commitments
Some carriers ask for volume commitments (we'll take X cars from you next month). The forecast gives the GM a defensible answer.
Cash Flow
Finance multiplies the forecast by average RO severity and expected close rate. The output is an AR-adjusted revenue projection. At PE-backed MSOs, this is a weekly or monthly conversation.
Parts Procurement
If the forecast shows 20% more volume next month, parts ordering and supplier conversations adjust now, not after the spike.
Intervention
If the forecast is dropping, leadership intervenes: marketing, DRP outreach, referral programs. The forecast becomes a trigger for action, not just a number.
How to Deploy It
Start Simple
Your first forecast should be a Prophet model running on 18+ months of daily deliveries by shop. One notebook, two afternoons. Get the baseline.
Measure the Error
Run the forecast in parallel with whatever you're currently doing for a month. Compare. Report the error honestly.
Add Features Incrementally
Once the baseline is trusted, add WIP, intake pipeline, weather. Measure whether each addition improves error. Stop when it stops improving.
Productionize
Wrap the model in a scheduled job. Output to your warehouse. Build dashboards. Don't leave the forecast in a notebook only the data person can run.
Feedback Loop
Track forecast vs actual weekly. When the forecast is wrong by more than 15% on a given shop, investigate. Usually the shop had an unusual event—flag it and retrain.
Why Most MSOs Don't Have This
Three reasons:
- No warehouse. Can't train a forecast on CCC exports pulled monthly.
- No one owns it. Forecasting is neither a pure ops function nor a pure finance function. It falls in the gap.
- The perfect-or-nothing trap. Leadership refuses to trust a 15%-error forecast because they're comparing it to a nonexistent perfect forecast. But they're currently running on a 30%-error gut feel.
All three are solvable. The first requires warehouse investment. The second requires someone to own the metric. The third requires framing: "our current forecast has 30% error; this one has 15%. Let's use the better one and keep improving."
The Honest Bottom Line
A good vehicles-out forecast is worth more than any single AI feature in collision. It drives staffing, cash flow, carrier negotiations, and capital planning. And it's technically boring—classical time-series methods with modern ML on top, deployed well.
That's why no one sells it aggressively. It doesn't demo well. But it pays for itself in the first quarter of deployment.