My five real records finished +190.00. I built four disclosed models to test what happens when that short result meets the current official odds, clustered noise and the same uneven stake pattern over a much longer run.
The line I will not blur
The real evidence behind this project is much smaller: five recorded bets in four CCTV Rush Hour rounds. I staked 560.00, received 750.00 back and finished those visible entries +190.00. Two selections won, three lost, and the final 200.00 Over bet returned 720.00.
That is a useful case study, but it is a terrible probability estimate. Two wins from five produces a 40% point estimate with a 95% interval stretching from roughly 11.8% to 76.9%. Almost any long-run story can hide inside a range that wide.
The five entries are not even five independent game rounds. Two different selections share round #RHO301769 and therefore share one settled traffic count. I keep both entries in the financial ledger because both stakes were real, but I do not pretend they provide two independent observations for estimating a win probability.
The official base probabilities put the expected return of that visible stake pattern at 504.00 from 560.00 staked, or −56.00 net. My actual result was +190.00 because the final 720.00 return supplied 96% of everything returned in the sample; without it, the same displayed ledger ends −330.00. Replaying the four-round pattern under the base model still has about a 20.675% chance of finishing at least +190.00. That is a good result, but not rare enough to demonstrate an edge.
So I kept the real records as the behavioural starting point—market choices, stake changes and realised returns—then separated them from the synthetic experiment. Every chart below is generated by code from a displayed model version and fixed seed. Reloading the page gives the same result.
The charts are synthetic outputs, not screenshots of 10,000 wagers, not a claim of long-run profit and not a prediction of the next camera count.
Primary-source inputs checked 21 August 2026
The current 155.io rules list a 90% base RTP on every Rush Hour market, with an effective range of 90–92.16% after eligible Lightning boosts. That is lower than the 91.5–93.5% figure shown on an older provider marketing surface and repeated in earlier parts of this site. For this experiment, I use the current detailed rules.
| Market | Gross payout | Base win probability | Base RTP |
|---|---|---|---|
| Range | 2.25x | 40% | 90% |
| Under | 3x | 30% | 90% |
| Over | 3.6x | 25% | 90% |
| Exact — either boundary | 18x | 5% | 90% |
| Exact Low / Exact High | 36x | 2.5% | 90% |
The camera count and the betting outcome are not the same layer. The provider's current specification says a committed seed draws the winning market, then the boundaries are placed around the AI-measured camera count. That means I do not use a Poisson traffic forecast to manufacture an edge.
Sources: 155.io Game Rules, Game Payouts & RTP and the official Rush Hour product page. I model settled base rounds only; refunds and Lightning boosts are excluded unless explicitly stated.
Four questions, four assumptions
I normalise each observed outcome to a one-unit stake, then repeatedly resample the five realised return multiples: 3x, zero, zero, zero and 3.6x. This deliberately treats my tiny sample as if it were the world. It is a diagnostic trap, not a forecast.
I alternate Under at 3x and Over at 3.6x, use one unit per round and set win probability to 90% divided by the multiplier. This is the clean reference model.
I add persistent low, neutral and high regimes plus ±3 percentage points of round noise while preserving a 90% marginal expectation. It tests streak sensitivity; it does not reconstruct the provider's actual seed system.
I apply one common outcome path to a flat 112-unit stake and my repeated 10/150/50/150/200 shape. Both wager exactly 1.12 million units, so only variance changes.
Fixed-seed output
The five-record bootstrap climbs because it keeps recycling two wins from a winner-heavy sample. The official baseline moves towards its expected −1,000-unit result. That contrast is the point of the experiment.
| 10,000-round model | Realised RTP | Net | Max drawdown | Longest loss run |
|---|---|---|---|---|
| Five-record bootstrap | 134.12% | +3,411.80 | 31.00 | 22 |
| Official independent baseline | 90.43% | −957.40 | 960.00 | 26 |
| Clustered noise sensitivity | 89.63% | −1,037.20 | 1,071.80 | 28 |
Simulated, not observed. Each line uses a fixed seed; the bootstrap is intentionally overfit to five real records.
A model can be perfectly reproducible and still answer the wrong question. Model 1 faithfully repeats my short sample and produces nonsense as a long-run forecast. Model 2 uses a defensible probability structure and finishes close to the theoretical loss.
Sensitivity, not a secret traffic system
Independent Bernoulli rounds are the clean baseline. Real players, however, experience results as streaks. To stress that experience, Model 3 moves between persistent low, neutral and high probability regimes, then adds uniform noise from −3 to +3 percentage points.
The regime model still targets the same 90% marginal return. It changes the path, not the long-run edge. In this fixed run, 100-round block RTP becomes more volatile: a 17.30-percentage-point standard deviation versus 14.03 in the independent baseline. Maximum drawdown grows from 960.00 to 1,071.80, and the longest losing run grows from 26 to 28.
The purple path is hypothetical. It demonstrates streak risk while preserving the same unconditional expectation.
Same edge, different ride
At one unit per round, every base market has the same expected 10,000-round net: −1,000 units. What changes is dispersion. Range concentrates outcomes tightly; Exact Boundary creates a much wider tail where a rare late hit can temporarily leave a run positive.
| Market | 10k net SD | Approx. 95% net interval | Chance of positive net |
|---|---|---|---|
| Range 2.25x | 110.23 | −1,216 to −784 | 9.00 × 10−18% |
| Under 3x | 137.48 | −1,270 to −730 | 2.86 × 10−11% |
| Over 3.6x | 155.88 | −1,306 to −694 | 1.18 × 10−8% |
| Exact 18x | 392.30 | −1,756 to −226 | 0.603% |
| Exact Low / Exact High 36x | 562.05 | −2,080 to +116 | 4.082% |
Analytic standard deviation for 10,000 flat one-unit base bets. A wider bar means more variance, not improved expected return.
The part I can control
My documented amounts were 10, 150, 50, 150 and 200. To isolate that shape, I repeated it and compared it with a flat 112-unit stake on exactly the same Under/Over outcomes. Both paths wager 1,120,000 units across 10,000 rounds, so both have an expected net of −112,000.
| Stake policy | Total wagered | Fixed-seed net | Realised RTP | Analytic net SD |
|---|---|---|---|---|
| Flat 112 each round | 1,120,000 | −89,488 | 92.01% | 16,461 |
| Repeated 10/150/50/150/200 | 1,120,000 | −76,450 | 93.17% | 19,453 |
The uneven pattern happens to finish 13,038 units better on this seed. It is still the riskier policy. Its analytic standard deviation is 18.2% higher, and its 95% outcome band is wider. The model did not discover a staking edge; it produced one luckier path.
Same official outcome path and equal turnover. The difference between the lines is variance and timing, not a change in house edge.
What I learned
My five records ended positively because one 3.6x return carried almost the whole result. When I bootstrap those records, the model confidently invents a 134% game. When I replace that tiny sample with current official probabilities, the positive slope disappears.
The noise model adds rougher streaks but does not remove the edge. The stake test moves wins and losses into different places but does not improve expectation. The high-multiplier markets create wider positive tails but also deeper negative ones.
My conclusion is not “never model a game.” It is “model the correct layer.” Use real records to inspect behaviour and interface evidence. Use current rules to model expected return. Use sensitivity scenarios to test uncertainty. Do not let five outcomes estimate a probability they cannot support.
For my next documented session, I would keep the stake range fixed and write the market choice before each result. That would not beat the 90% base return, but it would produce a cleaner comparison between observed play and the synthetic model.
No. I played the five documented bets in my gameplay report. This page contains synthetic settled rounds generated from disclosed assumptions.
Because it repeatedly resamples only five observed results, including two wins. It demonstrates small-sample overfitting and must not be read as a forecast.
No. The clustered model is a hypothetical stress test. Current provider rules say the committed seed draws the winning market and the boundaries are then fitted around the AI count.
No. With the same turnover and probabilities, stake shape changes variance and drawdown, not expected return.
