How accurate is it?
We put 832 real battle reports through the simulator, exactly as the reports show them, and compared its prediction with what really happened. They run from early-game armies all the way to T12 Fire Crystal 10, including some of the strongest accounts in the game.
788 of 832 real battles.
The favourite won 614 of 620 of those fights.
Real losses landed inside the predicted 80% range for 566 of 697 sides.
Only 3 of 756 player reports went a way it gave less than a 1-in-100 chance.
The 832 fights are 756 reports from players and 76 test cases published by another community project. The win-chance table and “Real surprises” use the player reports only, and the losses figures count each side of a fight separately: 756 sides that survived and 756 that were wiped out.
Gaia Heart is a PvE mode where the loss error is much wider, so the “inside the range” figure leaves it out. It has its own line in the table below: 33 of its 59 sides landed inside the range.
How far to trust the win chance
These are 755 real battle reports from players, put through the simulator. Each row groups the fights by the chance the simulator gave its favourite, and shows how often that favourite really won.
The other 1 of the 756 had no favourite (neither side won half of the simulated battles), so it isn't in this table.
| Simulator gives its favourite | Fights | Favourite won | Said Happened |
|---|---|---|---|
| 50–70%A close fight | 63 | 35 (56%) | |
| 70–90% | 72 | 63 (88%) | |
| 90–99% | 95 | 92 (97%) | |
| 99% or more | 525 | 522 (99.4%) |
A close call really is close: the favourite won about six in ten of the fights called 50–70%. Run it again with more battles before you march on one of those.
Losses come as a range
The game has luck in it, so the same two armies don't lose the same troops every time. The simulator runs your fight hundreds of times and shows the typical loss with an 80% range: eight in ten of its battles ended inside it.
For the real reports, the actual losses landed inside that range for 566 of 697 sides that weren't wiped out (81%). When a side was wiped out for real, the simulator's range included a wipe for 750 of 756 of them.
Each simulated battle also allows for what the simulator cannot know: its picture of a hero may be a little off, and a report does not show skill levels or which rally joiners counted. That allowance is sized on real reports, which is why the range holds about eight real results in ten.
- Won lost 3,693
- Lost lost 5,000
- Lost lost 5,000
- Won lost 3,062
- Won lost 3,104
- Won lost 2,853
One side won four and lost two, and its losses in the wins swung by 840 troops. That is the game's own luck, and it's why a single number would mislead you.
Results by type of fight
"Winner picked" counts the fights where the side the simulator favoured won in the report. "Typical loss error" is how far the predicted total losses usually are from the report: ±20% means usually within 20%. "Inside the range" counts sides that weren't wiped out.
| Type of fight | Fights | Winner picked | Typical loss error | Losses inside the range |
|---|---|---|---|---|
| Solo fights and duelsAttacks on a city or tile, Alliance Championship, Expedition Arena | 532 | 509 (96%) | ±16% | 81% |
| Rallies and garrisons | 131 | 116 (89%) | ±27% | 82% |
| Gaia Heart (heroes, PvE) | 59 | 58 (98%) | ±34% | 56% |
| Labyrinth (no heroes) | 28 | 26 (93%) | ±13% | 79% |
| The same duel fought six timesShows the game's own randomness | 6 | 4 (67%) | ±25% | 5 of 6 |
| Outside test cases, no heroesPublished by a separate community project | 76 | 75 (99%) | ±2% | – |
| Outside threshold tests18 of the 76 above, set up to sit right at the tipping point | (18) | 18 (100%) | ±0% | – |
| All fightsThe range column leaves out the outside test cases and Gaia Heart | 832 | 788 (95%) | ±2–34% by type | 81% |
Tested at every stage of the game
The reports cover small early-game armies, mid-game rallies, and T12 Fire Crystal 10 fights between some of the strongest accounts there are. Here they are by the highest attack bonus on either side.
| Strongest attack bonus in the fight | Fights | Winner picked |
|---|---|---|
| Under 1,000%Early game, T6 to T10 | 100 | 98 (98%) |
| 1,000% to 2,000% | 13 | 12 (92%) |
| 2,000% to 3,000% | 342 | 328 (96%) |
| 3,000% to 4,000% | 235 | 219 (93%) |
| 4,000% and aboveThe top of the game, T12 Fire Crystal 10 | 142 | 131 (92%) |
Built from the game's rules, not tuned to these reports
Since version 13 no strength in the engine is tuned to who won these reports or what they lost: the battle rules come from controlled test fights, and every skill works at the strength the game's own text states. One thing is sized on reports: the width of the range. 420 of the duels were always kept out of that tuning. On those 420 it picked the winner in 403 (96%), and their real losses landed inside the range 82% of the time.
The 76 outside test cases are a different matter: the battle rules come from projects that worked them out on controlled fights like these, so the engine matching them closely (±2%) is to be expected.
Where it is weakest
- Rallies are hard: more players, more heroes and more joiner skills, so the loss error is wide (±27%). Gaia Heart is the widest (±34%).
- Fights with no heroes play out almost the same every time in the simulator, so they get almost no range. Use the typical loss there (±2% to ±13%).
- It was very sure (95% or more) and wrong in 5 of the 832 fights. Each one is being looked into.
- In Gaia Heart the range holds only a little over half of the real results, not eight in ten, so treat its losses as a rough guide. Finding out why is the current work.
Every fight was run 500 times with the engine the site uses today. Reports players share with the accuracy data are checked before they join this test set.