Backtesting

The Whale.BackTest module provides a replay engine for evaluating betting strategies against historical outcome data. It applies the full Whale stack (sizing, bankroll, risk, policy, guards) on each simulated decision point and collects performance metrics.

Defining a strategy

A strategy is a function from market state to a list of candidate bets:

import Whale.BackTest exposing (Strategy, Candidate)
import Whale.Types exposing (Probability, Odds)

let my_strategy : Strategy =
  fun state ->
    state.opportunities
    |> List.filter (fun o -> o.edge > 0.02)
    |> List.map (fun o -> Candidate.new o.prob o.odds)

The backtest engine calls this function at each time step, feeds candidates through the sizing and guard pipeline, and records outcomes.

Running a backtest

import Whale.BackTest exposing (run, Config)

let config = Config.new
  { initial_bankroll = Money.usd 10_000.00
  , sizing = Sizing.fractional_kelly 0.5
  , policies = my_policies
  , guards = my_guards
  }

let results = BackTest.run config my_strategy historical_data

historical_data is a time-ordered sequence of market snapshots, each containing the opportunities visible at that point and the realized outcomes.

Results

BackTest.run returns a BackTest.Results record:

Field	Type	Meaning
final_bankroll	Money	Ending balance
peak_bankroll	Money	High-water mark reached
max_drawdown	Float	Largest peak-to-trough decline as a fraction
total_bets	Int	Number of bets placed
win_rate	Float	Fraction of bets that returned a profit
total_return	Float	(final - initial) / initial
sharpe	Float	Annualized Sharpe ratio of per-period returns
log_growth	Float	Realized geometric growth rate

Comparing strategies

Run multiple configurations over the same historical data and compare on the metrics that matter to your use case:

let results_full = BackTest.run config_full strategy data
let results_half = BackTest.run config_half strategy data

let better = if results_half.sharpe > results_full.sharpe
  then "half Kelly"
  else "full Kelly"

The backtesting surface is deterministic: identical inputs produce identical outputs. For Monte Carlo exploration (shuffled outcome order, synthetic data), generate the input sequences upstream and feed them through BackTest.run as separate historical datasets.