Backtesting option strategies

What a backtest answers

A payoff chart shows what a strategy does at expiry for every price; a backtest shows what it did on the prices that actually happened. Enter the same structure at regular intervals, exit by a rule, and add up the results. It tests a rule ("sell a 30-day strangle every week, hold to expiry"), not a single trade.

Reading the numbers

• Win rate: the share of trades that made money. On its own it misleads: short-option strategies win often and lose big.
• Total and average P&L: what the rule made, after costs.
• Maximum drawdown: the largest fall of cumulative P&L from a peak to a later trough. It is the pain you would have had to sit through, and a better guide to sizing than the average.
• P&L relative to maximum loss: each trade's result divided by what it could have lost, for structures whose loss is capped. It compares trades of different sizes on the same scale.
• Costs: the spreads crossed and fees paid, entry and exit. On options they are often the difference between a winning and a losing rule.

A worked example

A rule sells a short strangle 20 times. Sixteen trades keep most of their premium, +$900 each on average; four are caught by a large move, −$4,500 each on average.

• Win rate: 16 / 20 = 80%.
• Total P&L: 16 × $900 − 4 × $4,500 = $14,400 − $18,000 = −$3,600, an average of −$180 a trade.
• If two losing trades come back to back after a run of gains, the drawdown is at least $9,000, ten winning trades' worth.

An 80% win rate and a losing rule: this is the pattern behind probability of profit versus expected value.

The biases

• Look-ahead bias: using information that wasn't available at the time, such as choosing the strike with the day's closing price, or filtering on the volatility that followed.
• Survivorship bias: testing only what still exists, such as the coins and venues that survived, or the strikes that ended up liquid.
• Model prices: pricing history from a fitted volatility surface instead of recorded quotes. The model fills every gap, including the moments when nobody would have traded with you.
• Ignoring costs: testing at mid prices, which no one trades at.
• Overfitting: trying many rules and keeping the best. The best of fifty rules on one year of data is mostly luck.
• Too short a history: a year without a crash says little about a strategy whose risk is the crash.

The backtest in BOSS

BOSS's backtest replays a strategy over full option-chain snapshots that its ingest process records. At each entry it builds the structure from that snapshot exactly as a strategy page builds it live, enters at execution (asks for bought legs, bids for sold ones, fees included), and exits by closing at execution in a later snapshot or, held to expiry, at intrinsic value on the index recorded just before expiry. It uses no model: a trade whose legs had no size on the side they trade, or that can't be found at exit, is skipped and counted, not priced. Its summary shows the trades, wins, total and average P&L, best and worst trade, maximum drawdown, costs, and the average P&L relative to maximum loss.

The same strengths are its limits. Its history starts when recording started; it doesn't test calendars and diagonals; and settlement on the last recorded index approximates the venue's official delivery price. A deployment can also run without it.