Price trends methodology

Every figure comes from this site's own price tracking. We check retailer listings for gaming laptops on a repeating schedule, store what we see, and derive the published figures from that record. There is no third-party price feed and no survey.

Only laptops we actually track are in scope. A laptop that has never been listed here has never been measured here, so the figures describe the part of the market this site covers rather than the market as a whole.

No. Amazon is excluded from every figure on these pages, matching how the rest of the site treats Amazon in price charts and record lows.

Amazon prices move on a different rhythm from other retailers and are not comparable with them day to day, so including them would change what the index measures rather than widen it.

Each region is measured, published and read entirely on its own, in its own currency. Nothing is converted and nothing is merged across regions.

A converted figure would carry an exchange rate we did not measure, on a date we would have to choose, and it would move whenever the rate moved even though no laptop had changed price. Keeping regions apart keeps every number something we can point at a listing for.

A listing counts on a given day when it is one of ours to measure: the laptop is part of the tracked set for that region, the retailer is not Amazon, the listing was active that day, and the listing has been confirmed in stock at least once. A listing we have never seen in stock has no price history worth reading.

Only genuine observed price updates count. Markers an operator added by hand to describe the period before tracking began are excluded from every trends figure.

A price at or below zero cannot occur, and if one ever did the whole configuration-month would be dropped rather than published.

These are list prices. Every figure is the price shown on the retailer's listing, before any coupon or promotional code. Some retailers offer codes we track but cannot value — the code is recorded, the discount it gives is not — so netting them out would mean guessing. A shopper using a code may pay less than any figure here.

Discount depth compares a listing's sale price against the regular price the retailer itself publishes beside it.

That regular price is self-declared by the retailer and we do not verify it. A retailer that inflates its regular price will show a deeper discount here than a shopper really gets. Read discount depth as 'how big a saving retailers are claiming', not 'how much you are really saving'.

The cheapest-price figure is the one an error would distort most, because it is a single extreme rather than an average, so it carries its own guard.

A candidate for cheapest price is rejected if it sits more than 35% below the bottom of the market for its group, if it rests on fewer than 20 qualifying days that month, or if its own daily price bounced around too much to be a settled price. When a candidate is rejected the next-cheapest is tested instead.

Averages and medians are left alone: they are robust to a single bad reading by construction, and 'cleaning' them would mean deciding by hand which observations to believe.

Prices are recorded when they change, not once per day, so a day between two checks has no reading of its own. For those days the last known price is carried forward, but only for a limited number of days and only while we still have grounds to believe the laptop was in stock.

The carry limit depends on how densely that period was recorded: 5 days for the well-recorded present, 30 days for the sparser early period. Past the limit the day simply has no price and is dropped from the calculation.

Where a laptop is listed by several retailers, the day's figure for that laptop is taken from the eligible listings for that day; it is not an average of retailers that were not selling it.

A single configuration only gets a monthly price if it was really present that month: at least 15 qualifying days and at least 50% of the month's days.

Those configuration months become family figures, and the family figures become the published market prices for the group. The typical price is the middle of that spread, the good-value price is its lower quartile, and the cheapest price is the lowest survivor of the outlier guard.

A month is frozen when it is published. The figures for a closed month do not move afterwards unless a correction is issued, and a correction is announced on the page rather than applied quietly.

A matched comparison only counts laptops present in both periods being compared. It answers 'did the same laptops get cheaper', which is the question a price index is for.

That is deliberately different from comparing this month's average against last month's. If three cheap models launch and two expensive ones are discontinued, the average falls without any laptop changing price. A matched comparison is unmoved by that; the current market section on the trends page is the place where launches and discontinuations legitimately show up.

A laptop entering or leaving the tracked set mid-period is not a price change and is excluded from the matched pair rather than treated as one.

The index chains together month-on-month matched changes, so each link compares like with like even as the set of laptops on the market turns over.

When there is no valid link for a month - too few laptops present in both months, for example - the chain cannot bridge the gap. Rather than invent a link, the index restarts, and the chart draws the break as a break instead of a line. Two levels either side of a break are not comparable and the page never joins them.

A family is a laptop model line; a configuration is a specific specification of it. Configurations are combined into a family figure first, and families are then combined into the published figure.

This is a deliberate choice rather than a technicality. Counting every configuration equally would let one model that ships in fifteen variants outvote five models that ship in one, and the published price would then track that manufacturer's product-line strategy rather than the market. Every published figure states both counts, so the sample behind it is always visible.

Different figures need different amounts of evidence, so each one has its own minimum sample and each is withheld on its own terms. A missing figure always shows the reason in plain language rather than a blank.

Typical price needs at least 8 families and 12 configurations. Good-value price needs 12 families and 20 configurations, because a lower quartile drawn from a handful of laptops is not really a quartile. Cheapest price needs 8 families and 12 configurations plus the outlier guard. A matched change needs 8 matched families and 12 matched configurations. Availability needs 12 configurations and 200 configuration-days. Discount depth needs 8 families and 20 configuration-days with a published regular price.

Counts are never withheld. A count is its own sample size, and zero is a real answer: seeing 'no typical price' beside 'seven families' tells you the group was one family short, whereas two blanks tell you nothing.

Before any figure in a month is published, the month as a whole has to be well enough covered: at least 95% of eligible configuration-days backed by publishable evidence, at least 90% of configurations meeting their own monthly coverage rule, and at least 95% coverage on the month's first and last day so a boundary month cannot be half-measured.

A month that fails is not published anywhere, in any form. It is not published greyed out, not charted differently and not published with a caveat.

That last point is the reason the rule is strict. A caveat travels badly: a screenshot, a quoted sentence or a spreadsheet cell all lose the footnote, and the number keeps circulating without it. A number that needs a caveat to be honest should not be published.

A price only counts on a day when we have grounds to believe the laptop was actually buyable. The strongest evidence is a reading taken that day. Next strongest is a price carried forward across a window that stock records confirm on both sides.

Weaker than that, a carried price rests only on the absence of a record saying the laptop went out of stock - and during a period when little was being recorded, that absence is not evidence. Those days are dropped from every published figure.

This is why the archive starts when it does rather than reaching back to the beginning of the site's history. The earlier period is honestly reconstructable but not honestly publishable, so it is not published.

Current prices count every laptop that is in stock at a tracked retailer, at whatever price it is listed at today. A laptop does not have to be discounted to be counted.

The question these figures answer is 'what does a laptop with this GPU cost today', so a laptop selling at its full price is as much a part of the answer as one that has been reduced. If a discount were required, the whole section would disappear whenever retailers stopped running sales - which is exactly when knowing the going rate matters most.

What a laptop does need is to be buyable and recently seen: in stock at a tracked retailer, with a price we have observed within the last 5 days. A laptop we have not checked recently enough is left out rather than published as though it were still on the shelf.

Because they answer different questions, and both answers can be right at once.

The monthly publication asks whether the same laptops got cheaper. The current market section asks what a laptop with a given GPU costs today. When a cheaper generation launches, the current market price falls even if no existing laptop was discounted; when an expensive model is discontinued, the same thing happens in reverse.

They also cover different time spans. The monthly publication is a closed, frozen month. The current market is measured as of the last refresh, over a trailing 30-day window, and moves whenever prices do.

A published month is never edited in place. A correction is published as a new revision at its own permanent address, and the page shows what changed, when and why.

The revision it replaces stays permanently readable at its own address, so anyone who cited the original can still see exactly what they cited.

Yes. Every published month offers a CSV and a JSON download of exactly the figures on the page, for that region, that month and that revision.

The files are generated once when the month is published and served unchanged, so a download taken today and the same download taken next year are the same bytes. They accept no filters, no column selection and no ranges.

Attribution is on the page rather than inside the files, so the CSV opens cleanly in a spreadsheet or a data tool without a licence banner in the first row.

The files use the published field names rather than the labels on the page, and those names never change: what the page calls the good-value price is value_market_price in every CSV, JSON payload and citation, because they were frozen when the first month was published and people have quoted them since.

No public API response, embedded page state, or downloadable export contains a product-level record. The only laptops the trends archive names are the small number of hand-picked highlights rendered on the page itself, each of which already has a public product page on this site. Every machine-readable request returns aggregates conforming to the published aggregate export schema.

The published cells are region, GPU tier and month. There is no free-form filter that could narrow a cell down to a single product, and no cell is published below its minimum sample - the sample rules are themselves what stops a figure describing one laptop.

Yes, with attribution, under the Creative Commons Attribution 4.0 licence. Each page carries a ready-made citation with the region, period, revision, publication date, methodology version and permanent addresses.

Please cite the revision as well as the month. It is what makes a quoted figure checkable if the month is ever corrected.

This page describes methodology version 1.0.0. Every published month records the methodology version it was produced under, and that version is part of its citation.

Changing how a figure is calculated means a new methodology version. Figures produced under different versions are labelled as such rather than silently mixed.