Coin flip guide
Eagle Coin Flip: what a simulation shows and hides
Eagle is no more than the label on this page. The page runs a seeded computer simulation of the coin flip format and compares it with exact theory.
Why simulate a game you can solve exactly?
Every average for a coin toss can be written down with pencil and paper. A simulation adds something different: it shows what individual sessions look like, with all their noise. Averages describe the centre; a simulated session shows the ride.
The simulation
Below are eight sessions of 150 one-unit flips each, at a 1.98x payout. The random numbers come from Python's pseudo-random generator started from seed 77. Change the seed and every row changes. None of this is a record of real play.
| Session | Rounds won | Net (units) | Lowest point | Longest losing run |
|---|---|---|---|---|
| 1 | 73 | -5.46 | -7.76 | 5 |
| 2 | 70 | -11.40 | -20.08 | 8 |
| 3 | 75 | -1.50 | -6.40 | 8 |
| 4 | 79 | 6.42 | -1.12 | 6 |
| 5 | 87 | 22.26 | -0.04 | 5 |
| 6 | 70 | -11.40 | -12.38 | 7 |
| 7 | 78 | 4.44 | -3.34 | 6 |
| 8 | 80 | 8.40 | -2.24 | 5 |
Seeded pseudo-random simulation (Python random.Random, seed 77), 8 sessions of 150 one-unit rounds. Average session result 1.47 units. This is a simulation, not a record of real play; a different seed gives different numbers.
Reading it against theory
- Rounds won. The expected count is 75 out of 150, with a standard deviation of about 6.1. Rows between roughly 69 and 81 are typical.
- Net. Theory says -1.50 units per session on average, with a standard deviation near 12.1. Rows far from the mean in both directions are expected.
- Lowest point. Even sessions that finish ahead often dip well below zero first. That is the drawdown a player must sit through.
- Longest losing run. The exact expected longest losing run in 150 flips is about 6.6.
When does the simulation converge?
With enough rounds a simulation homes in on the exact value. Using seed 77 again, the mean net per flip over 1,000 flips is -0.0001; over 100,000 flips it is -0.0120. The exact answer is -0.0100. The error at n flips is roughly one divided by the square root of n, so 100,000 flips still leave a few thousandths of noise.
Limits of simulation
A pseudo-random generator is fine for teaching but is not how a hash-verified service makes results. For the service side, see the hash check page. A simulation can also mislead if the seed is cherry-picked to tell a story, which is why the seed is printed above.
Could a different seed show all eight sessions winning?
Yes, it is possible, just unlikely. That is exactly why one simulation proves nothing about the long run, and why the exact expected value matters.
Why 150 flips per session?
It is long enough to show drawdowns and streaks, and short enough that luck still outweighs the edge, which is how many real sessions feel.