Research methodology: The $500 Experiment

Version 1 · method version igs-methods-v1 · betting rules version 2026.1

This document is pre-registered: before the first wager of a season, its SHA-256 fingerprint and the time of registration are recorded and published. Anyone can download this file and check that its fingerprint matches. Once registered, it is not edited; any change is published as a dated deviation (section 11).

1. Purpose

To measure, in public, what happens when ordinary people wager imaginary credits at real sportsbook prices over a season. The project's name asks a question. This study is designed to answer it fairly, including if the answer is unexpected.

2. Design

An open, voluntary, observational study. Participants are not randomly assigned to anything.
Each season, every enrolled participant receives 500 credits with no cash value. There are no deposits, purchases, withdrawals or prizes.
Every participant may bet. Not betting keeps all 500 credits, so 500 is the line every bettor is measured against; no separate non-betting group is recruited.
Wagers are moneylines (two-way, or three-way for soccer), point spreads, totals, and parlays of 2 to 10 legs from different games, at the price shown on the board: for each market, the first of DraftKings, FanDuel and BetMGM that currently offers it. Prices are locked at placement and graded on official final scores under the published betting rules.

2a. Feedback shown to participants

From the first wager of the season, every participant sees the same automatic feedback on their private dashboard, and may choose to show it on a share link:

Explanations. Each price is shown with its implied probability; each slip shows its fair chance (margin removed by multiplicative normalization), the return a margin-free book would pay, and the house's cut. These are descriptions of the price, not predictions.
Lab notes (badges). Findings about the participant's own settled bets: Beat the close (at least 20 picks with a closing price, 55% or more of them with a positive closing edge as defined in section 5), Beat the coin (at least 20 settled bets ahead of the random-bettor benchmark), Parlay alchemist (a won parlay with 4 or more winning legs), Well calibrated (at least 10 answered singles, stated confidence within 5 points of the win rate), Even Steven (at least 20 settled bets within 2% of break-even), Variance survivor (a losing streak of 6 or more), The confidence gap (at least 10 answered singles, stated confidence 20 or more points above the win rate), Parlay graveyard (5 or more settled parlays, none won), Closing-line casualty (at least 20 picks with a closing price, 40% or fewer of them with a positive closing edge), Bust (under one credit left).
Classification. One label from the participant's own numbers, never a rank against others: Bust; Unclassified (fewer than 20 settled bets); Wizard, provisional (at least 30 picks with a closing price and a mean closing edge of 2% or more); Lucky (ahead without that edge); Coin flipper (return on stakes within 5% of zero); Margin payer (the rest).

No feedback rewards the number, speed or timing of bets, and the thresholds above are minimum sample sizes, not targets. Feedback may still change behavior (section 9); because it is identical for everyone and present from the start, it is part of the condition being studied, not a treatment.

3. Participants and eligibility

Adults (18 or older, self-attested) with one account each.
Taking part in research is a condition of enrollment, and every participant is included in published research under a subject number. Participants who withdraw from research, leave the season or delete their account are removed from future analyses.
Excluded from analysis: staff accounts, and enrollments the team marks as test accounts or as duplicates of another participant (for example, identical betting histories placed within minutes of each other). Each marked exclusion records its reason and date. The number excluded is published with the results.

4. Hypotheses

Primary (confirmatory):

H1. Among bettors with at least 20 graded wagers by the end of the season, more than half finish with a negative net result.
H2. Participants' stated confidence in single bets is poorly calibrated: bets placed with higher stated confidence win less often than the stated confidence implies.
H3. Bettors' average closing edge is not greater than zero (they do not, on average, beat the closing line).

Secondary (exploratory, reported as such): whether results differ by sport, by bet type (singles versus parlays; moneylines, spreads, totals), by self-reported experience and knowledge, by number of wagers, whether early winners stay ahead, and whether stated interest in real-money betting changes between the baseline questionnaire and an end-of-season follow-up.

5. Outcome measures

All monetary amounts are integer centi-credits; the measures below are computed from the ledger and settled wagers exactly as stored.

Measure	Definition
Graded wager	A wager that WON, LOST or PUSHED. VOID wagers are excluded from performance measures
Net result	Σ (amount credited − stake) over graded wagers
Return on stakes	Net result ÷ Σ stake over graded wagers
Bankroll return	(equity − 500) ÷ 500, where equity = available credits + stakes of open wagers
Standing	Ahead, behind or even according to the sign of the net result
Bust	Equity below 1.00 credit (the minimum stake)
Closing edge	Per leg: decimal odds taken × fair probability at close − 1. The fair probability comes from the last quote from the same book and market fetched before the event's start, with the margin removed by multiplicative normalization. Legs whose line differs at close are excluded, and the share of legs covered is reported
Random-bettor benchmark	Per wager: the mean, over every combination of one side per leg at the locked prices, of the combination's return at the same stake (a leg's factor is that side's decimal odds if it won, 1 if it pushed or was void, 0 if it lost; a combination's return is stake × the product of its factors). A combination whose potential return exceeds the payout cap would have been refused at placement, so it returns its stake. This is the exact expected return of choosing sides uniformly at random, with no simulation
Calibration	For singles with a stated confidence: by confidence level, the number of bets, the share won (pushes and voids excluded), and the mean implied probability of the price taken

6. Comparison groups

The starting 500. Not betting keeps all 500 credits, so the share of bettors whose equity is below 500 is the share who would have been better off not betting. Participants who never bet, and bettors who stopped (no wager in 14 days), are reported as separate groups.
Random bettor. Same events, markets, prices and stakes as each real wager; isolates selection from the bookmaker's margin and from stake sizing.
Closing line. The market's final assessment before the event; the standard measure of betting skill that does not depend on short-run luck.

7. Analysis plan

H1. One-sided exact binomial test that the share of eligible bettors with a negative net result exceeds 0.5, with a 95% Wilson interval. Also reported, descriptively, for all bettors and by wager-count band (1–4, 5–19, 20+).
H2. For singles with a stated confidence: the Brier score of stated confidence against outcomes, the calibration slope from a logistic regression of the outcome on stated confidence, and the calibration table. H2 is supported if the slope is below 1 and the 95% confidence interval excludes 1. Standard errors are clustered by participant.
H3. Mean closing edge across legs with a comparable close, with a 95% confidence interval from a participant-clustered bootstrap (10,000 resamples). H3 is supported if the upper bound of the one-sided interval is at or below zero.
Multiple comparisons. Holm correction across H1–H3. All secondary analyses are labeled exploratory and not used to claim confirmation.
Timing. The confirmatory analysis uses data at the close of the season, after the last settlement and correction window. Interim figures published during the season are descriptive.
Missing data. Optional questionnaire answers are not imputed; analyses that use them report how many participants answered.

8. Publication and privacy

Published aggregates include a sample size. Any figure describing fewer than 10 participants is withheld; a wager-level figure also needs at least 10 wagers (or picks), from at least 10 different participants.
When a breakdown withholds a group, the next-smallest groups are withheld too until the withheld groups cover at least 10 participants, so no group can be recovered by subtraction from a total. Group sizes below 10 are shown only as "fewer than 10".
The day-by-day share of bettors ahead is published to the nearest whole percent, without daily counts. A live dashboard can still change when one person's result changes; this residual risk is accepted for figures that never name or single out anyone.
Individual histories are never published. A participant may share a one-page summary of their own results through a revocable link that shows their subject number, or a nickname only if they choose.
The published aggregates are recorded daily, with the method version, so every published figure can be reproduced.

9. Known limitations

Imaginary credits change behavior: people may bet more boldly than with real money.
Participants self-select; results describe this population, not all bettors, and comparisons are not causal.
One season is short; over a few months most of a participant's net result is variance, which is why closing edge is a primary measure.
Prices come from three US books via a data provider and are refreshed hourly and shortly before each event; between refreshes, prices can be older than at a live sportsbook.
Stake and payout limits (1 to 100 credits per wager, 50,000 credits maximum return) constrain behavior and are applied to everyone equally.
Participants see explanations and findings about their own results (section 2a). Seeing the margin, or a "Bust" note, may change how they bet afterwards. Every participant sees the same feedback from the start, so the study describes bettors who are shown the arithmetic, not bettors kept in the dark.

10. Data and grading integrity

Credits are stored as integers and the ledger is append-only. Every graded wager records the exact quote and result version it was graded on. Settlement waits 30 minutes after a final score; later corrections re-grade affected wagers and are posted as visible corrections. The full betting rules are published at /rules.

11. Deviations

Any departure from this plan, including a change to a definition, an exclusion, or an analysis, is published with its date, what changed, and why, before the affected results are reported.

12. Registration

Registration records, for the season: the SHA-256 fingerprint of this file (exact bytes, UTF-8, LF line endings) and the time of registration. The database accepts a registration only before the season's first wager, and the record cannot be altered afterwards. The time of the first wager is shown next to the registration.