New vehicle registrations run at roughly five percent of the vehicles already on the road, so this year's parts demand is set by sales that happened six to thirteen years ago and are already visible in registration data. The demand unit is the application rather than the part number, which makes the fitment join to the parc the single most valuable and least maintained asset in the business. Supersession and quality tiers split one real demand stream across several part numbers, and pooling theory says a branch network holds several times the stock a central pool would, which is the price of answering the counter in four minutes.
A fitter rings the counter at ten past eight with a 2014 hatchback on the ramp and a seized caliper. He wants a part number, a price, and a yes or no on availability. If the yes takes longer than the time it takes him to ring the factor two streets away, the job goes there, and so does the brake fluid, the pads and the discs he was going to buy at the same time.
That phone call is the unit of demand in this business. Behind it sits a catalogue running to several hundred thousand lines, most of which sell in ones and twos with long gaps, and a demand driver that has almost nothing to do with how many cars the manufacturers sold last quarter. Most planning setups arriving from a consumer goods or manufacturing background are built for a different shape of problem entirely, and the mismatch shows up first as an availability number that nobody trusts.
New registrations tell you very little about this year
S&P Global Mobility put the average age of light vehicles in operation in the United States at 12.8 years in May 2025, across a fleet of 289 million, with new registrations above 16 million in 2024 and a scrappage rate of about 4.5 percent. ACEA has the EU passenger car fleet at an average of 12.3 years. Both numbers say the same thing about planning horizon: annual additions are around five percent of the population, so a strong or weak year for new car sales moves the composition of the parc by a fraction of a percent.
The arithmetic that follows is the useful part. A vehicle generates very little independent aftermarket demand in its first three or four years, because warranty and franchised servicing absorb it. It becomes a serious consumer of parts through the independent channel somewhere around year five and stays there until it leaves the road. So the parc that pays your wages this year was registered between roughly six and thirteen years ago, and those registration counts are already recorded. You are forecasting against a population that has already been observed.
A cohort also decays at a rate you can measure. At a fleet-wide scrappage of 4.5 percent a year, a cohort loses half its units in about fifteen years, and older cohorts leave faster than the average implies. The consequence is that a part fitted only to a platform built between 2008 and 2013 has a demand curve that has already turned over, and no amount of tuning a time series model on that part's own sales will tell you whether the decline is a share problem or a parc problem. The cohort and hazard machinery for turning an installed population into a parts forecast is the subject of a separate post and the mechanics are the same. What differs here is that the population belongs to nobody you can call.
The join between your catalogue and the parc
Your demand unit is a part number. The parc is counted by vehicle configuration. Everything depends on the table that connects them, and in the United States that table is standardised: the Auto Care Association maintains the Vehicle Configuration Database, with more than 60,000 year, make and model combinations and monthly updates, alongside ACES for fitment and PIES for product attributes. TecDoc plays the equivalent role across much of Europe. These are the industry's shared vocabulary for saying which part goes on which car.
Treat that join as a planning input rather than as catalogue housekeeping and the forecast changes shape. Work an example with illustrative numbers. A front pad set carries 780 applications. Joined to registration data for your trading area, those applications add up to 1.9 million vehicles in the age band you actually sell into. Front pads get replaced on average once every four years of use, which puts the annual replacement pool at around 475,000 sets. Independent workshops served by factors handle perhaps 40 percent of that work, giving 190,000 sets addressable through your channel. At a six percent share of that channel you would expect around 11,400 sets a year across your branches.
If you sold 8,900, the gap is a commercial conversation about coverage, price position and which workshops you are losing, and it arrives with a size attached. A forecast built from the item's own history would have told you 8,900 plus a trend, and would never have raised the question.
The same join breaks in ways that look like demand movements and are not. A supplier adds 60 applications to a part in a data release and your parc-based expectation jumps overnight with no vehicle having changed. An application removed after a supersession quietly deletes demand you are still selling. Version the fitment data and diff it between releases, because an unversioned join means every forecast movement has two possible causes and no way to separate them.
Four versions of the same physical part
The catalogue carries the same function several times over. There is an original equipment part, a premium aftermarket brand, a mid tier, a value line, and often your own label. Functionally these compete for one job on one car, and a customer who cannot get the tier he asked for will usually take the one above or below it rather than wait.
That makes the forecasting unit larger than the part number and smaller than the category. The workable level is the interchange group: every part number that would satisfy the same application. Forecast demand at that level, where the history is denser and the substitution is invisible, then split it across tiers using observed mix and price position. Planning each tier as an independent series produces the familiar pattern where the value line stocks out, the premium line sits, and the total looks fine.
Range breadth across tiers is a separate decision with its own economics, and the general treatment of how wide a catalogue should be is covered under assortment rationalisation. What belongs here is the sequencing point: decide the tier ladder first, then plan the group, then allocate. Doing it the other way gives you five thin forecasts instead of one usable one.
Supersession turns one stream into three
A part number gets replaced by a successor, which is later replaced again. The demand is one continuous stream. The history sits in three places, and each fragment reads as an item in decline followed by an item with no history.
The damage is specific. New item logic assigns the successor a launch profile it does not need. The predecessor gets flagged for obsolescence review while it is still selling. Any statistical method fitted to the fragments sees dying series everywhere. The fix is unglamorous: maintain a supersession chain as a first class object, roll history forward onto the current number, and plan the chain rather than its links. Chains also branch, where one part is replaced by two, and merge, where two converge on one, so the mapping needs weights rather than a simple pointer.
This is worth doing before any modelling work, because every downstream method inherits the fragmentation. It is also the piece most often left to whoever last touched the catalogue.
The counter promise decides where stock sits
Pooling theory is clear about what a branch network costs. Eppen showed in 1979 that consolidating locations with independent, identically distributed demand reduces expected holding and shortage cost in proportion to the square root of the number of locations, formalising what Maister had described in 1976 as the square root law. Thirty branches each holding two of an item that sells ninety a year nationally means sixty units on hand against ninety of annual demand. A single pool serving the same demand would need a fraction of that.
Nobody in this industry gets to take the saving, because the promise at the counter is measured in minutes. What you can do is decide, per item, whether the promise has to be met from the shelf in front of the fitter or from a hub on a two hour shuttle. Both of those are a yes at the counter, and only one of them carries thirty positions.
The rule that works uses order lines rather than units. An item earns a branch position when it appears on enough separate customer orders at that branch to make a local unit likely to move, which in practice starts somewhere around one demand occasion a month. Items below that line sit centrally with a defined delivery window, and the counter quotes the window rather than the stock. Note the caveat on the square root result, which Maister and later reappraisals were explicit about: it assumes demand at each location is independent and similarly distributed, and correlated regional demand or a very lumpy item weakens the saving considerably. Use it to size the prize, then check it against your own branch level history.
The demand you refuse never reaches the forecast
Counter systems record invoices. The call where you had nothing to sell leaves no trace, and the fitter does not ring back to tell you what he bought instead. Every item you are short of therefore reports a demand history depressed by exactly the amount you failed to serve, and the replenishment calculation built on that history orders less, which produces more shortfall.
Aftermarket distribution has this problem worse than most channels because the substitution happens outside your business within four minutes. The recovery methods for censored history are the subject of a separate post. The prerequisite belongs here, and it is a capture problem rather than a statistical one. Most counter and trade systems can already log a lookup that ended without a line: the part enquired for, the vehicle, the branch, the time, and whether it was quoted from stock, from the hub, or refused. Turning that log on costs a configuration change and starts building the only record that will ever tell you what your true demand looks like.
The limit
You buy parc data rather than measuring it, and its quality varies enormously by market. In territories with clean registration records you can get vehicles in operation by model, engine variant and postal district. Elsewhere you get national counts by make and year, and the join to a fitment table keyed on engine code degrades into guesswork. When the parc data is coarse, a parc-based forecast is a category level sanity check rather than an item level plan, and it is honest to use it that way.
The bottom of the catalogue also resists all of this. For the tail that sells once or twice a year in a territory, no demand model earns its keep. That is a range decision about what you are willing to be the shop that has, and it belongs to a commercial owner with a stated cost of holding it, rather than to a planner with a forecast.
The one structural change ahead that history cannot inform is parc composition. Battery electric vehicles consume no oil filters, no spark plugs and far less brake friction material because of regenerative braking, while carrying suspension, steering and climate parts much like anything else. Your exposure is arithmetic you can run today from the same registration data, by tagging the applications behind each part with a powertrain type and reading off what share of your gross margin sits on parts that only combustion vehicles need. That number is a fact about your catalogue rather than a forecast about the market.
This week, take your top 200 part numbers by margin, join them through the fitment table to vehicles in operation by model year for your own trading area, and see how many are sitting on a parc that has already started to shrink.
More on this in industry on the DemandLabs platform.
