transparent by design

How Carconomics turns prices into decisions.

A field guide to every important number on the search and stats pages: what it measures, which observations feed it, where it helps, and where it cannot.

methodology.ts

source        = real trip quotes
refresh       = about every 30 minutes
active_window = ORD Jul 11, 2026 → Sep 8, 2026 · LAX Jul 11, 2026 → Sep 8, 2026
last_update   = Jul 11, 12:22 PM CT
output        = evidence, not a booking guarantee

◉ LIVE latest successful refresh⟲ HISTORY comparisons across time◇ PLANNED structure only, not live

00 / reference

The system contract#

The site compresses a large grid of messy rental prices into decisions while keeping the evidence visible.

for renters

Find the trip, not merely the car.

Compare exact totals across dates and lengths, expose discount cliffs, and show why a result looks unusually valuable.

for hosts

Read the market without guessing.

See current price bands, timing patterns, reputation cohorts, and where longer-trip discounts create unintended bargains.

decision-contract.txt

raw quote grid
  → normalize the same trip dimensions
  → compare like with like
  → expose the strongest signal
  → link back to the bookable listing

Rule: descriptive evidence stays separate from prediction. A current price curve says what is listed now; a history signal says how identical trips changed. Neither is certainty about tomorrow.

01 / reference

Real trip quotes#

A marketing daily rate is not enough. Carconomics works from the trip total for a specific car, pickup date, and duration.

quoted_total

data: real, bookable rental prices near ORD and LAX
window: refreshed about every 30 minutes
how: each car is checked across every pickup date in the active market window and each trip length from 1 to 14 days

The number shown is the trip total before taxes, not an advertised “from” rate. Checkout-only protection, tax, and optional charges must be confirmed before booking.

daily_price

data: live prices (◉)
window: the selected car + pickup + trip length
how: quoted trip total ÷ number of rental days

Daily price is a comparison unit, not a separately bookable price. Every duration is quoted independently because host discounts make the curve uneven.

+N days saves $X

data: live prices (◉)
window: same car and pickup, across longer trip lengths
how: find the cheapest longer duration whose full total is below the selected duration's total

A weekly or long-trip discount can overshoot. This flag shows the exact longer option instead of assuming every added day costs more.

quote.math

quote_key = market + listing + pickup_date + rental_days
per_day   = quoted_total / rental_days
free_day  = longer_total < selected_total
book_link = same listing + exact compared dates

02 / reference

Search modes#

Every mode resolves to the same date-and-duration combinations on the client and server, so a shared URL means the same plan everywhere.

Flexible

Starts from one target date, expands pickup flexibility, and enforces a hard minimum trip length with an optional maximum.

Date range

Tests every trip inside the chosen range that meets the minimum length.

Mass Search

Tests the stored pickup window across all supported lengths, or one selected length.

around-date.length

minimum = shortest acceptable trip
maximum = selected cap or "No max"
No max  = search through the current 14-day data ceiling

example: minimum 3d + maximum 6d → 3d, 4d, 5d, 6d

result_order

data: live prices (◉)
window: all quotes matching the active plan and filters
how: sort by trip total, daily price, year, rating, or price relative to fair value; deterministic fields break ties

Filters define the eligible set first. Highlighted picks and sort order then come from that same set.

03 / reference

Fair value is a comparison, not an appraisal#

The model estimates a normal daily price for the same kind of car and trip, then compares the real quote with that reference.

fair-value.model

expected_daily_price = f(
  market, model, vehicle_age, rating_presence, rating,
  completed_trips, all_star_status, trip_length,
  days_until_pickup, pickup_weekday
)

value_gap = 1 - actual_daily_price / expected_daily_price

fair_value

data: live prices (◉)
window: refit after each successful refresh
how: a regularized pricing model learns current relationships between vehicle, reputation, market, and trip-timing characteristics

“22% under fair” means the quote is 22% below the model’s expectation for those observed characteristics. It is not an appraisal or a checkout guarantee.

Steal

data: live prices (◉)
window: latest fitted model and quote set
how: requires a material saving, a listing-weighted held-out residual cutoff, global validation, and enough listings in that market and model

The label is selective. If R², held-out absolute error, sample support, or segment support fails, the badge is not served; fair-value sorting still degrades safely.

Best_value

data: live prices (◉)
window: the active search result set
how: actual daily price ÷ modeled fair daily price, lowest ratio first

This favors unusually discounted quotes instead of automatically favoring the smallest or oldest car. It remains a price signal, not an overall quality score.

04 / reference

Distance stays independent#

A cheap car farther away can still be a genuine price bargain. Travel inconvenience is a separate personal tradeoff, so it is not hidden inside fair value.

price value does not equal location convenience

Fair value is distance-neutral. The model adjusts for the broad market, but it does not reward or penalize a quote based on miles from an airport or from you.

distance.contract

candidate_set = selected market or future user radius
price_value   = actual price vs comparable price
distance      = shown and filtered independently
user_choice   = whether the savings justify the travel

Today: where available, cards show distance from the market reference point. Planned: users will be able to provide a location, choose a radius, and see user-relative distance. That radius is not live yet.

05 / reference

Mass Search#

Mass Search asks the expensive but useful question: what are the strongest opportunities anywhere in the current date-and-length surface?

mass-search.plan

matched_quotes =
  selected markets
  × every pickup date in the stored window
  × every supported duration
  × every eligible car

response = ranked page + match count + explainable picks

Great_deal

data: live prices (◉)
window: all quotes matched by the Mass Search request
how: a daily price materially below the middle of that request's matched prices

This is a quick within-search price flag. It is distinct from Steal, which uses the fitted fair-value model.

result_page

data: live prices (◉)
window: one stable slice of the full ranked match set
how: each refresh precomputes all 450 Mass filter sets, next-local-day variants, and five exact top-400 orders; the server decodes only the requested cards

The total match count can be much larger than the visible cards. The first 400 are exact for that filter and sort; narrow the plan to explore deeper instead of treating this as a bulk-export API.

Performance contract: growth should increase indexed data, not browser work. Large searches must stay bounded, paginated, deterministic, and able to retain the last valid dataset if a new publication fails.

06 / reference

Freshness, missing rows, and carries#

Rental inventory can disappear briefly even when a listing is valid. A short carry window prevents one incomplete refresh from blanking the site.

freshness.state

observed now        → publish current quote
missing once        → carry last known quote
missing twice       → carry last known quote
missing three times → remove quote
new window day      → remove expired pickup date

◉ live vs ⟲ history

data: each panel identifies its evidence class
window: ◉ = latest successful refresh · ⟲ = observations accumulated across days
how: live panels use the current union; history panels compare bounded observations over time

A carried quote can trail the source by roughly two refresh intervals and is labeled “last-seen price” on cards. Host Desk excludes carried targets and peers. Always confirm final price and availability.

checked_at

data: timestamp of the dataset's latest successful publication
window: shown in the global status badge
how: the prior valid dataset remains available if a refresh or publication fails

The badge describes the dataset as a whole. It does not guarantee every individual row was observed during that exact refresh because short carries are allowed.

07 / reference

Stats provenance#

The stats page is a market terminal: each panel should answer a decision question and disclose whether it uses the latest surface or repeated observations.

signal	source	calculation	use it for
price heatmap	◉ LIVE	typical $/day by pickup date × length	finding inexpensive trip windows
lead-time curve	◉ LIVE	current quotes by days until pickup	reading today’s price surface
model / weekday / reputation	◉ LIVE	current quote medians by cohort	comparing observable segments
price drivers	◉ LIVE	controlled fitted-model effects	separating correlated traits
marginal day / free day	◉ LIVE	fresh same-car and pickup pairs across lengths	spotting discount cliffs
market trend	⟲ HISTORY	daily market-local medians	tracking broad direction
do prices drop as pickup nears?	⟲ HISTORY	same trip followed across days	observing repricing behavior
recent price drops	⟲ HISTORY	same quote becomes at least 5% cheaper	finding fresh changes

Important: a median is quote-weighted unless a panel explicitly says it is listing-weighted. “After controls” is a model association, not proof that a host badge or vehicle trait caused a price difference.

08 / reference

What the data cannot prove#

Useful analysis gets stronger when its boundaries are explicit. These limits apply even when a signal looks compelling.

01
Not the checkout total. Quotes are before tax; protection and checkout-only options are outside the dataset.
02
Not availability demand. A listing disappearing can mean many things. It is not called a booking without direct evidence.
03
Not a causal experiment. Controlled price effects reduce obvious mix differences but do not prove causation.
04
Not review-text analysis. Current quality signals use aggregate rating, completed listing trips, and All-Star status. Public pages expose some reviews, but automated/commercial collection stays off unless Turo grants written permission for that use.
05
Not a prediction by default. A current lead-time curve compares different future trips today. Only repeated same-trip history measures repricing.
06
Not permanent. Price and availability can change at any moment. The linked listing remains the final source.

09 / reference

Bounded by design#

Every stored dataset has a replacement rule or retention limit. Growth should add coverage without creating an unbounded data pile.

current quote surfacewhole-replaced

Fixed-width quotes plus bounded indexes; publication stops safely at 20 MiB.

price changes48 hours

Recent drops stay actionable and then expire.

refresh detail7 days

Cadence stays inspectable without accumulating forever.

market indices + drift180 days

Runway for calibrated time comparisons.

daily summaries400 days

A bounded year-over-year analytical window.

daily research snapshots400 days

Full-resolution analysis history, separate from serving.

growth.guardrail

new_store_allowed = retention_rule && size_bound && owner
publication       = validate first, point current last
failure           = preserve prior valid generation
compatibility     = old reader + new reader during migration

10 / reference

Future vehicle categories#

The first expansion beyond Tesla listings — the Luxury group — is live; specification rankings still require trustworthy trim-level data first.

◉ LIVE The Luxury group is live. Its curated membership is make-level for Porsche, BMW, Mercedes-Benz, Audi, and Lexus, and model-level elsewhere: Range Rover and Range Rover Sport, plus the Escalade family and CT5-V/CT4-V — model years 2022 and newer only. It runs the same real-quote methodology described above on its own separate data store, refreshes about every two hours, and is filtered by make where the Tesla group is filtered by model. Groups are never blended: every price, band, and badge is computed within one group at a time.

vehicle-catalog.schema

identity = make + model + year + trim
specs    = MSRP + acceleration + power + range + seats + cargo
quality  = exact | probable | unknown
rule     = uncertain inputs never become a confident ranking

listing-metadata.plan

permission_gate = written source license
claimed_fields  = trim + seats + range + features + delivery
verified_specs  = separate licensed catalog or VIN match
raw_reviews     = never retained by default
retention       = overwrite current + delete after 90d unseen

fastest / $

Verified performance divided by exact daily rental price.

MSRP / rental $

Inflation-adjusted original vehicle value per exact daily rental cost.

horsepower / $

Verified output per rental dollar for the selected trip.

range value

Usable driving range relative to the daily quote.

luxury bargain

High vehicle-value percentile at a low rental-price percentile.

family value

Seats, cargo, and range under an explicit daily budget.

Every future leaderboard will run after location/radius, dates, duration, and budget filters. It will show its inputs and formula—never collapse price, speed, distance, and quality into one unexplained score.

Find the trip, not merely the car.

Read the market without guessing.

const quoted_total

const daily_price

const +N days saves $X

Flexible

Date range

Mass Search

const result_order

const fair_value

const Steal

const Best_value