Bid-ask bounce is the reason tick-level signals look better than they are
When trades alternate between hitting the bid and lifting the ask, the last traded price flips back and forth even if the true value stays the same. This bounce produces a negative autocorrelation in tick returns that looks like mean reversion.
A strategy that sells after an up-tick and buys after a down-tick will appear profitable in a backtest using last prices, because it is simply capturing the bounce. In reality, to exploit it, the trader would have to cross the spread on every trade, paying exactly the amount the bounce provides.
The illusion is widespread in tick-based testing. Results look wonderful until execution costs are properly modeled, at which point the edge disappears.
To avoid the trap, test with separate bid and ask series, buying at the ask and selling at the bid. Alternatively, use the mid price for signals but charge the full spread on every fill.
Any tick-level system that does not survive that treatment was only harvesting noise.
Has a tick-level backtest ever fallen apart once bid and ask were separated?
Bid-ask bounce is the reason tick-level signals look better than they are
-
Amanhasnoname
- Posts: 2060
- Joined: Fri Oct 30, 2026 9:45 am
Re: Bid-ask bounce is the reason tick-level signals look better than they are
A tick-fade I wrote for EURUSD looked almost too clean. Using last price it showed sixty-four percent winners and about four tenths of a pip expectancy across eighty thousand simulated trades. The moment I forced buys at the ask and sales at the bid, expectancy dropped to minus one tenth and the win rate fell to forty-seven percent. The whole edge had been the bounce between the two sides.
I also tried signaling on the mid price and charging a full spread on every fill. That version was slightly worse, because the two-tenths spread I had modeled was kinder than the live book at the London open. Waiting for a two-tick mid move before entering still lost after costs.
I scrapped the system. Any tick idea I test now starts on separate bid and ask series, and if it only works on last traded price I do not take it any further.
I also tried signaling on the mid price and charging a full spread on every fill. That version was slightly worse, because the two-tenths spread I had modeled was kinder than the live book at the London open. Waiting for a two-tick mid move before entering still lost after costs.
I scrapped the system. Any tick idea I test now starts on separate bid and ask series, and if it only works on last traded price I do not take it any further.