For twenty years, the used-car business ran on three books: Kelley Blue Book, Black Book, and Manheim Market Report. Every appraisal, every trade-in, every wholesale bid started with a number pulled from one of them. That number was always a look at the recent past — the average of what similar cars sold for last week, last month, or last quarter.
In late 2025 the category shifted. The largest players stopped selling rear-view valuations and started selling forecasts. If you buy or sell used cars — as a shopper, a flipper, or a dealer — the old playbook is broken, and the reason is worth understanding.
What changed in 2025
Three announcements told the story.
In December 2025, CarGurus launched PriceVantage, a predictive pricing tool that pairs real-time shopper demand signals with machine-learning models to project where a specific vehicle's price should sit over the coming weeks — not where it sat last month. It was CarGurus' first move from "here's the market average" to "here's the market next."
One month earlier, CDK Global argued that predictive AI is now "crucial" for used-car inventory decisions and released an update to its stocking recommendation engine that trains on rolling market data instead of monthly snapshots. The framing mattered: CDK positioned prediction as table stakes, not as an add-on.
Then in December, Manheim expanded the MMR — the wholesale price benchmark every US dealer references — to include broader "range" data and richer digital signals across their unified marketplace. That's a lagging-index vendor acknowledging that a single-point-in-time average isn't enough anymore.
Three data points don't make a trend on their own. But they do make the same trend: the incumbents who owned the "what is this car worth today" question are all racing to answer "what will this car be worth in 30, 60, 90 days." The competitive floor moved forward in time.
Why lagging books mislead in a volatile market
The three books average recently observed transactions. In a stable market that works — the last four weeks are a decent proxy for the next four weeks. In a volatile market, it doesn't.
Consider what "recent" means when a book updates weekly. If wholesale prices on a segment moved 4% in the last two weeks, a book that averages the prior 30 days is already stale by roughly half that move. A dealer who bought against that book paid the average of a market that no longer exists.
The last three years show how often the market gets volatile. Semiconductor shortages inverted the normal depreciation curve on used trucks and SUVs in 2021 and 2022. EV incentives and price cuts on new inventory pushed used EV values down sharply in 2023 and 2024. Interest-rate cycles moved monthly payments faster than sticker prices, which means dealer floor-plan math shifts even when the vehicle itself doesn't. On the /market-trends dashboard we track wholesale index moves against consumer-facing prices — the two curves are rarely aligned week-to-week.
Concretely: our own weekly data shows the used compact-SUV segment softening 2–3% across Q2 2026 while dealer asks on the same VINs held flat. A dealer stocking against the book saw margin compress into the trade. A dealer with a forward view either held off or discounted acquisition. That's the practical cost of using a rear-view number to make a forward-looking decision.
What "AI forecasting" actually means
The phrase "AI forecasting" gets used to describe everything from a simple trailing-average to a large neural network. It's worth being precise about what a real forecast is.
At CarCast we build time-series forecasting models. They take a long history of weekly prices for a specific vehicle segment — a 2023 Toyota Camry LE, a 2020 Honda Civic EX — and learn how that price series behaves. They condition on macroeconomic signals that empirically move used-car prices: the used-car component of the Consumer Price Index, the Manheim wholesale index, regional CPI variation, gas prices, and rate-cycle indicators. The model then projects the next 60 days.
Three properties distinguish this from a book:
Confidence bands, not single points. Every forecast comes with a probability distribution. The P50 is the median expected price. The P10 and P90 mark the 10th and 90th percentiles — you know how much variance the model expects around the median. On a segment where recent history has been stable, the bands are tight and you can act with confidence. On a segment where the model has seen wide swings, the bands widen and you know to hedge. A single-number book can't tell you which regime you're in.
Weekly retraining. Fresh transactions get ingested every week, and the models retrain on the new window. That's what closes the "recent doesn't equal future" gap. A book updates its history; a forecasting model updates its expectations.
Segment-level, not VIN-level. A forecast on "2023 Toyota Camry LE" is a statement about the segment. A specific VIN with 84,000 miles and a scuffed bumper still needs a human condition adjustment on top. That trade-off is deliberate: segment-level forecasts have enough weekly transaction volume to be statistically meaningful. VIN-level "forecasts" don't. Anyone selling you a VIN-specific 60-day forecast is either doing very heavy smoothing on very sparse data or extrapolating a segment forecast without saying so. We publish our methodology and accuracy publicly on the /accuracy page for exactly this reason — a forecast without a track record is a guess with better marketing.
The right mental model: a forecast is a probability distribution over near-future prices for a segment, refreshed weekly, honest about its confidence.
What AI can't do
Being clear about the limits is more useful than another benefit bullet.
No condition or mileage adjustment at the segment level. The forecast is for a representative unit in that segment. A specific car with high mileage, salvage history, or immaculate condition needs a human adjustment on top. The book's edge here is precisely the granular condition/mileage math — a good workflow uses both.
No prediction of one-off shocks. Recall announcements, sudden tariff changes, manufacturer-side incentive shifts, and new-vehicle production disruptions do move used prices, but they show up in the model only after they show up in the data. A forecast issued the week before a shock will miss it. Once the shock is in the price series — typically within two to four weeks — the model reflects it. Treat forecasts as base-rate reasoning, not oracle.
No prescription. A rising forecast is a signal to consider acting, not a command. There are legitimate reasons to buy against a softening forecast (an inventory hole in your lot, a specific customer request, an unusually clean example) and reasons to wait against a rising one (cash-flow constraints, better opportunities elsewhere). Forecasts inform decisions; they don't make them.
Not financial advice. Vehicle purchases involve financing, insurance, taxes, and personal circumstances well outside a price model's scope. Our forecasts are informational analytics for a specific segment's expected transaction price. If you're making a large financial decision, the forecast is one input among several.
Human judgment stays in the loop. What changes is the quality of the base rate the judgment sits on top of.
How to use forecasts if you're a buyer
If you're shopping for a used car in the next 90 days, the workflow is simple.
Look up the segment you're considering on CarCast — for example, /vehicles/model/2020-honda-civic — and check the 60-day forecast direction. Three cases:
- Rising forecast. Prices are expected to climb. If you were planning to buy in the next month anyway, moving sooner likely saves money. Bring the forecast printout to the dealer as a talking point.
- Softening forecast. Prices are expected to fall. If your timeline is flexible, waiting three to six weeks typically captures some of the move. Set a price alert on the segment so you know when it hits your target.
- Flat forecast with tight bands. The market is stable. Timing gives you little edge — decide on non-price factors like condition, mileage, and warranty coverage.
Cross-check the forecast against inventory. A softening forecast plus rising inventory count is a stronger buy-wait signal than a softening forecast alone.
How to use forecasts if you're a dealer
For an independent used-car dealer or wholesaler, the highest-leverage use of a forecast is acquisition, not appraisal. The book is fine for appraising a specific VIN in front of you. The forecast tells you which segments to acquire from auction next week.
The workflow: run your dashboard's segments and sort by 60-day forecast direction. Segments with rising forecasts and tight confidence bands are your target inventory for the coming month — you're buying below where the model expects to sell. Segments with softening forecasts are ones to underweight, or to move faster if you already hold them. On $30,000 vehicles, a 3–5% forecast edge is $900–$1,500 per unit — real margin on volume.
Two guardrails. Never trade a forecast against a segment you don't move — the model's precision matters less than your turn rate. And always sanity-check the confidence band; a wide-band signal is directional information, not a green light. Full pricing and multi-user access lives on /pricing.
Try it yourself
CarCast publishes public 60-day forecasts across our full universe of tracked segments — every year/make/model/trim we have enough weekly transaction data for. Look up any vehicle you're considering, see the forecast direction and confidence bands, and decide whether to buy, wait, or hold.
Public forecasts are free, no signup required. If you're a dealer running acquisition against a lot, the Pro tier at $149/mo unlocks alerts, saved segments, and the Chrome extension for auction-day workflow. Wholesalers and multi-user teams should look at the /used-supercar-forecast hub and the Agency tier for volume access. MMR looks back. CarCast looks forward.