Case 10 · Simulation
Case 10 — the account that comes back from the dead
Scope: this case is about a simulation bug, not a market claim. It shows how a one-line ruin guard turns a wiped account into a survivor, and how to detect the shape statically. Reproduce with
auditkit/checks/bust_revival.py.
The shape
cap = 1.0
for r in returns:
cap *= (1 + lev * r)
if cap <= 0:
cap = 0.01 # <-- the account is revived at 1% and keeps trading
A leveraged path that loses more than the margin is over. Resetting the balance to 1% of the starting capital does two things at once:
- it makes the bust invisible in the equity curve (the account simply continues), and
- it lets later winners compound from a fresh stake, which the real account never had.
The same bug appears as a floor: cap = max(cap * (1 + r), 0.01).
What it did to one of our own engines
bt_withdraw_design.py (5x, 487 events, forward 24h returns rebuilt after the
column-semantics bug in case 8). With the revival floor kept, the withdrawal rules
looked like this (median total wealth, 216 year-block permutations):
| rule | with cap = 0.01 floor |
with "stop when ruined" |
|---|---|---|
| A no withdrawal | 0.2x | 0.00x (99.5% bust at K=1) |
| B withdraw 1x at 2x | 23.0x | 1.00x |
| C1 withdraw above 2x | 32.2x | 0.17x |
| C2 above 3x | 42.9x | 0.00x |
| C3 above 5x | 64.9x | 0.00x |
| C4 above 10x | 116.8x | 0.00x |
| D ladder 2/4/8 | 25.2x | 1.09x |
The left column is the bug. The right column is the same rules with a single change: a wiped position loses its margin and the account stops. At 5x with intrabar liquidation, K=1..5 busts in 42–99.5% of paths; only a 1/10-sized book (K=10) survives.
Detection
bust_revival parses with ast and flags two shapes:
REVIVAL_ASSIGN— an assignment to a small positive constant inside anifthat compares the same variable to zero.HIGHwhen the variable is multiplied (compounded) in the same function, otherwiseMEDIUM.REVIVAL_CLAMP—x = max(expr, 0.01). A floor of1e-9is a division guard and is deliberately not flagged; the cutoff is0.1.
On three of our own source trees (excluding virtualenvs) it found 36 files, 8 HIGH, 51 MEDIUM, one parse error. The HIGH files are the compounded engines; the MEDIUM ones are mostly per-trade floors that need a human look.
Why a linter and not a proof
A tiny floor is sometimes legitimate: price = max(price, 1e-9) guards a division.
The detector cannot tell intent. It points at the line, and the reviewer answers one
question: if this account were ruined here, should the next bar still trade?
If the answer is no, return 0 — do not hand it 1%.
Reproduce
python3 -m unittest tests.test_bust_revival -v
python3 - <<"PY"
from auditkit.checks import bust_revival
print(bust_revival.detect_file("path/to/engine.py"))
PY