exitkit — a catalogue of exit policies, with a backtesting.py adapter

[charlieyanhx]

I kept re-writing the same exit logic across strategies — a stop here, a time limit there, a volatility-scaled variant when the first two disagreed — so I packaged the family and wrote an adapter for backtesting.py.

exitkit — twenty-seven exit models in six families (stop-loss, take-profit, time-based, volatility, signal-reversal, convergence) behind one interface.

from backtesting import Backtest, Strategy
from exitkit import StopLossExitModel, FixedTimeExitModel
from exitkit.adapters.backtesting_py import ExitMixin

class SmaCross(ExitMixin, Strategy):
    exit_models = [StopLossExitModel(0.05), FixedTimeExitModel(24 * 30)]

    def next(self):
        self.apply_exits()          # close whatever the policy says to close
        if crossover(self.s1, self.s2) and not self.position:
            self.buy()

pip install exitkit[backtesting]

The point is comparison. Same entry rule, different exits, on the GOOG sample data that ships with this library:

exit policy           return %  trades  max DD %  Sharpe
none (hold)              326.1       1     -65.3    0.47
stop 2%                   47.6       8     -64.6    0.15
take profit 10%          187.5      10     -64.0    0.39
time limit 30d           208.3      31     -22.3    0.68
stop 5% + tp 10%         100.0      29     -28.3    0.49

One dataset and one entry rule, so it's an illustration rather than a finding — but swapping the policy is a one-line change, which is the part I wanted.

Two things I got wrong while building the adapter, in case they're useful to anyone doing something similar:

Holding time has to come off the bar clock. My first version derived it from time.time(), which is fine live and wrong in a replay — every 2010 position aged to now and fired every time-based exit on the first bar. There's a test asserting a ten-year holding limit doesn't fire on the GOOG data.

Two policies can fire on the same trade in one bar. The adapter tracks what it has already closed, so the second one is ignored rather than calling .close() on a closed trade.

It's deliberately narrow — it decides when to close and nothing else. No data feed, no order routing, no backtest loop; this library already does all of that better than I would. SignalOutput is a plain dataclass so it's a translation layer, not an adoption.

MIT, 141 tests, CI across 3.9–3.13. Feedback welcome, especially on the adapter's shape — it's the part most likely to be wrong.