Sensitivity to Cost Assumptions
A backtest's transaction cost model is an assumption, not a measurement — and a strategy whose profitability depends on that assumption being exactly right, rather than roughly right, is not a strategy to trust with real money.
Prerequisites: Sensitivity to Rebalance Frequency, Gross Versus Net-of-Cost Performance
A backtest assumes 5 basis points of round-trip transaction cost per trade and reports a net Sharpe ratio of 1.3. The number looks solid — until someone reruns the same backtest at 10 basis points, a perfectly reasonable estimate for a slightly less liquid universe or a slightly more urgent execution style, and net Sharpe drops to 0.4. The strategy wasn't lying at 5 basis points; the assumption itself was simply a guess, made before any real trading happened, and the strategy's entire viability turned out to hinge on that guess landing close to reality. Every backtest cost model is an estimate, never a measurement, and a result this sensitive to the estimate is a result that hasn't really been validated yet.
Why cost assumptions deserve their own sensitivity sweep
Transaction cost models bundle several separately uncertain pieces: the bid-ask spread paid to cross, market impact from the strategy's own trading (which depends on trade size relative to available liquidity, itself hard to estimate from historical data alone), commissions and fees, and any financing or borrow costs for short positions. Each of these is easy to get wrong by a meaningful margin, especially before a strategy has actually traded live and generated real fill data to calibrate against. Because these costs are subtracted from every single trade, an error in the cost assumption doesn't just shift the final number a little — it compounds across however many trades the strategy makes, so a strategy with high turnover is far more exposed to cost-assumption error than a low-turnover one, even if both use the same per-trade cost estimate.
Worked example: a breakeven cost sweep
A strategy generates 8% gross annualized return with 40× annual turnover. Net return at various assumed round-trip cost levels: at 2bps, net return is ; at 5bps, ; at 10bps, ; at 20bps, — the exact breakeven cost. Any realistic execution scenario above 20 basis points of round-trip cost makes this strategy unprofitable. If the research team's actual live execution desk typically achieves 12-18 basis points round-trip on similarly sized, similarly liquid trades, this strategy is uncomfortably close to its breakeven point and should be treated as marginal at best, not as the solidly profitable 8%-gross headline suggests.
What this means in practice
Every backtest report should state its breakeven transaction cost alongside its headline net performance number, and that breakeven cost should be compared honestly against what the execution desk actually expects to achieve for that specific universe and trade size — not against an optimistic textbook figure. A strategy with a large cushion between assumed cost and breakeven cost is robust to reasonable estimation error; a strategy sitting close to its breakeven line is one bad execution day, one liquidity-thinning market regime, or one slightly-too-optimistic initial assumption away from being unprofitable, regardless of how good the backtest's headline number looks.
Transaction cost assumptions in a backtest are estimates, not measurements, and should be swept across a plausible range — not fixed at one convenient value — to find the strategy's breakeven cost. A large cushion between assumed cost and breakeven cost signals robustness; a strategy sitting close to breakeven is fragile no matter how attractive its headline net Sharpe ratio looks.
Cost assumptions calibrated only on the strategy's own historical average trade size will understate costs once the strategy trades at real, potentially larger size — market impact grows faster than linearly with trade size relative to available liquidity, so a cost model that looked fine in backtesting can break down specifically once the strategy is scaled toward its intended capital.
Related concepts
Practice in interviews
Further reading
- Kissell, The Science of Algorithmic Trading and Portfolio Management, ch. 5
- de Prado, Advances in Financial Machine Learning, ch. 8