Forecast Methodology
How the forecast works
The price forecast on every rent page is deliberately simple and deliberately honest. Here is exactly what it does, and what it does not do.
The model: a quantile band with drift
For each GPU, provider, and region we keep a daily record of the hourly rental price. To forecast a future date we look at the last 90 days of those observations and do three things:
- Filter outliers. Any observation more than two standard deviations from the recent median is dropped first. Provider APIs occasionally report a glitch price or a transient $0 row, and those should not move the forecast.
- Estimate the centre. We start from the most recent price and add a small drift, the recent trend projected forward. This is the key honesty correction: a flat band would give the same number for 7 days out and 60 days out, which makes “pick a date” meaningless. The drift term makes a forward date actually move.
- Draw an 80% band. The 10th-to-90th-percentile spread of the recent window is centred on that drifted estimate, giving a low / central / high range rather than a single false-precision number.
The band widens past 30 days. Forecasting 60 days out is genuinely harder than 7, and the picture should say so.
Confidence labels
Every forecast carries a plain-language confidence label tied to how far out you are looking:
- High: up to 30 days out. The band is tight; provider rate cards rarely move this fast.
- Medium: 31 to 60 days out. The band noticeably widens; some providers may adjust pricing.
- Low: 61 to 90 days out. A ballpark only; rate-card changes are very plausible.
Confidence is degraded one level if the freshest price we have is more than 24 hours old. A forecast built on stale data should never present as High.
The 90-day cap
We hard-stop the forecast at 90 days. Ask for a date further out and the page says “beyond forecast range, check back closer to your date” rather than fabricating a band. Provider pricing is driven by business decisions, not a market we can extrapolate indefinitely, so a longer horizon would be guesswork dressed up as analysis.
When there isn't enough data
A forecast needs enough observations over a long enough span to mean anything. Until a SKU has accumulated a meaningful history, the page shows the current price with no forward band and explains that the forecast is still warming up. We would rather show nothing than show a confident-looking line drawn from three data points.
Provenance on every forecast
No forecast appears without its receipts. Each one can be expanded to show, in plain language:
- N observations: how many price points the forecast is built from.
- Date range: the first and last observation dates.
- Sources: which provider feeds the data came from.
- Last refresh: how recently we pulled fresh prices.
- Model: the method named above.
- Why this confidence: a one-line reason, for example “target date is 47 days out, so Medium”.
A forecast, not a prediction
This is a forecast, not a promise. It is a transparent, interpretable read on where prices have been heading, useful for deciding whether to rent today, wait a week, or wait a couple of months. It cannot know when a provider will rewrite its rate card overnight. We chose an auditable quantile-and-drift method over a black-box model precisely so you can see how every number was reached and judge it for yourself.