What every number on this site means, how it's calculated, and where it falls short. Fare Terminal only ever states what was observed and clearly labels anything inferred.
Airfare is not a single, stable price. The same route can be quoted differently by different sellers within minutes, seat inventory changes as fares sell, fare rules and included baggage vary by fare brand, and "the price" often depends on exactly which dates you search. A single quote tells you almost nothing about whether that price is unusual, expected, or a once-in-a-month opportunity. Fare Terminal exists to answer that question at the market level: given everything observed for a route recently, is the current price high, low, or normal — and how confident should you be in that read?
Every route tracked here is a market — an origin/destination airport pair, observed repeatedly over time under one of two search modes. A flexible market aggregates offers across a rolling departure window (currently 21–90 days out, 5–9night stays for round trips), which is the best signal for "is this route generally cheap right now." An exact market repeats the same specific depart/return dates over time, which answers a narrower question: is this trip currently priced well relative to its own history. They are tracked as separate series and are never averaged together — see compatible history below.
The current benchmark is the median of the 5 lowest valid, unique offers observed in the current window (median of fewer if fewer than 5 valid offers exist). We use a small low-price median rather than the single cheapest fare because one offer can be a data glitch, an expiring fare, or simply unavailable a minute later — a median of the cheapest handful is far more stable while still representing what a price-conscious shopper would actually see. "Valid" excludes expired offers and offers flagged as likely data anomalies (see data quality).
The from price is simply the single cheapest valid offer observed — the "starting at" number you'd see advertised elsewhere. It's shown alongside the benchmark rather than in place of it because the from price is more volatile (a single seat selling out changes it) and easier to game; the benchmark is the more trustworthy signal for "is this market cheap," while the from price answers "what's the best single deal right now."
Every derived snapshot records the methodology version that produced it (currently benchmark-v1). Percentile, fair value, and recommendation calculations only ever compare snapshots computed under the same methodology version — if the benchmark formula changes in the future, old snapshots are never silently mixed with new ones. This means a methodology change temporarily shrinks available history (and can trigger insufficient data states) rather than quietly comparing apples to oranges.
The percentile sentence ("cheaper than X% of observed history") is the share of this market's compatible historical benchmark prices that are strictly higher than today's. 100% means today is the cheapest this market has ever been observed; 0% means nothing in history was pricier than today. It requires no minimum history to compute, but reads as more meaningful the longer a market has been tracked.
The fair value band is a robust statistical range, not a "should cost" opinion: center = median of compatible historical benchmark prices; half-width = 1.5 × 1.4826 × the median absolute deviation (MAD) of that same history. MAD-based bands are used instead of a mean ± standard deviation because airfare history is skewed and has real outliers (fare sales, data glitches) that would distort a mean-based band; MAD is far less sensitive to those outliers. The band only renders once at least 15compatible historical snapshots exist — below that, the estimate isn't considered trustworthy and the UI says so explicitly rather than showing a shaky range.
Each market gets a composite score from five signed dimensions — historical value (based on percentile), 7-day momentum, departure-date urgency, offer-supply trend, and a volatility penalty — each clamped to a bounded range so no single factor can dominate. The label is read off the total score:
| Label | Score range |
|---|---|
| Buy | ≥ 3 |
| Lean buy | 1.5 – 2.99 |
| Neutral | -1.49 – 1.49 |
| Wait | ≤ -1.5 |
A market is instead marked insufficient data — with no buy/wait advice at all — whenever it fails any of three gates: fewer than 10 compatible historical snapshots, a data quality score below 0.35, or data staler than the freshness limit below. The UI states exactly which gate(s) failed rather than guessing.
The recommendation label (buy/lean buy/neutral/wait) is about the direction and strength of the signal; confidence (low/moderate/high) is a separate axis about how much to trust that signal. High confidence requires data quality ≥ 0.7, at least 30 compatible historical snapshots, and volatility ≤ 15%; moderate confidence relaxes those to 0.5, 15 snapshots, and 25% respectively. A strong "buy" score with low confidence is a real, if noisy, signal — not a contradiction.
Every event separates what was observed (a plain factual statement: a price moved, an offer count changed, a carrier appeared or disappeared from the cheapest set) from what is inferred (an interpretation of why it might matter), and every inference carries its own confidence level.
| Category | Event types |
|---|---|
| Price | Price drop, price increase, new historical low, low-fare set changed |
| Carrier | Carrier entered/left low set, possible carrier match |
| Fare product | Fare product appeared/disappeared |
| Volume | Offer count surge/contraction |
| Anomaly | Volatility spike, data anomaly |
Same-type events within 24hours of each other coalesce into one episode (the stored event's end time extends) rather than spamming the timeline with near-duplicate entries; a severity escalation breaks through that cooldown as a new event.
When two or more carriers' cheapest fares move the same direction within a short window (currently 6 hours) by at least 8%, Fare Terminal surfaces it as a possible carrier match — worded as "consistent with coordinated pricing but does not confirm it." This is deliberate: simultaneous price moves can reflect genuine competitive matching, shared underlying cost/demand signals, or plain coincidence, and the data available here cannot distinguish between those causes. The event is never labeled as confirmed competitive behavior.
Data is considered stale once it is more than 360 minutes (6.0 hours) old. Every price and timestamp on this site is labeled with when it was last updated, and a stale warning is shown explicitly rather than silently presenting old data as current.
Each snapshot gets a 0–1 data quality score: the unweighted average of an offer-count component (full credit at 8+ valid offers), a freshness component, and a cleanliness component (share of the input batch that wasn't expired or flagged as a likely anomaly). Low-quality snapshots are excluded from Market Pulse cards and can trigger insufficient-data recommendation states.
When configured against a real data source (TravelPayouts / Aviasales), Fare Terminal is reading cached, aggregated "cheapest price seen" observations sourced from real traveler searches — not a live GDS/NDC quote triggered on demand. Those cache entries can be up to roughly 48 hours old, carry no fare-brand/seat-count/booking-class detail, and their per-leg segment data is synthesized (not a verified itinerary) to fit this app's shared offer shape. Every such offer is flagged accordingly in its quality flags, and none of it should be read as "bookable right now at this price." See docs/PROVIDERS.md in the repository for the full adapter-level detail.
When this deployment is running against the TravelPayouts / Aviasales adapter, every price shown is a cached, aggregated "cheapest price seen" observation, sourced from real traveler searches — not a query this app triggers against an airline in real time. Per TravelPayouts' own documentation, those cache entries can be up to roughly 48 hours oldand are retained for 2–7 days; there is no guarantee the fare is still bookable at that price, or at all. A banner at the top of every page discloses this whenever it applies, and the "Check current availability" link on each market and offer takes you to the live booking site to verify before you act on anything shown here.
The adapter also cannot see real per-leg segment data for these offers — only one airline code, flight number, and total duration per result — so depart/arrive times and duration are reconstructed, not observed: round-trip flight time is split 50/50 between outbound and return (there's no way to recover the true split), and a missing duration falls back to a fixed placeholder. Offers built this way carry SYNTHETIC_SEGMENTS, ESTIMATED_LEG_SPLIT, and/or ESTIMATED_DURATIONquality flags, and the offers table marks their depart/arrive/duration cells "est." accordingly. Fare brand, booking class, and seat count are never known from this source either and always render "Not provided." See docs/PROVIDERS.md for the full adapter-level detail behind this section.
This deployment may be running entirely on synthetic, deterministic demo data (fictional carriers, real airport codes) rather than a live provider. Whenever that's the case, a persistent banner reading "Synthetic demo data. Not current airfare." is shown at the top of every page, and freshness is computed relative to the dataset's own anchor time rather than the real clock, so a demo deployment never falsely appears to have gone stale.
Nothing on Fare Terminal is a guarantee, a live quote, or booking advice. Prices change continuously and can differ from what's shown here by the time you check an airline or travel site directly. Recommendations are a statistical read of historical patterns, not a prediction — treat every label, percentile, and fair-value band as one input among many, always verify the actual price before booking, and see the about page for what this tool is and isn't.