A backtest on Orlin Minerals gives a Sharpe ratio of 2.1. The price file has one duplicated day and one bad print at ten times the true price that reverts the next day. What does cleaning do to the Sharpe, and how would you catch these errors automatically?

Cleaning takes the Sharpe from 2.1 to about 0.6. The bad print gives the strategy two giant winning days: long into the tenfold jump, then short into the fall back. Those two days make about 97% of the raw return. The duplicated day barely moves the number here but can shift every later signal by a day in other pipelines. Catch both automatically with a duplicate-date check and a spike-and-reversal filter before any backtest runs.

Plot it before computing anything. A tenfold price that returns to normal the next day is not a market move; it is a data error, usually a misplaced decimal or a wrong instrument code. Think of a bank statement showing Rs 1,00,000 credited on Monday and debited on Tuesday by mistake: your balance is right by Wednesday, but a spending plan built on Monday's balance is not. The duplicated row is quieter: the same close appears twice, so the file shows a day on which nothing happened.

Walk the strategy through the error. The day before the print Orlin fell, so the strategy is long. The price then jumps tenfold, a log return of about 2.30, and the next day it falls back by the same amount while the strategy, fading the jump, is short: two winning days, each a log return of about 2.3. The clean strategy earns about 10% over the whole six months; the raw file shows about 473%. Two fictional days are about 97% of the raw return. They also raise the volatility, which is why the Sharpe ratio reaches 2.1 and not something far larger.

Because its harm depends on the pipeline, not on this backtest. If returns and signals are ever joined by row position instead of by date, one extra row shifts every later signal by a day, and a strategy that looks one day into the future can show a spectacular Sharpe ratio for months. It also inflates the day count used to annualise volatility and hides a real missing day somewhere else, because vendors often duplicate the previous close when a day's data failed to arrive.

Build checks that run before any research code sees the data. Reject duplicate dates and flag runs of identical closes. Flag any day whose return is more than, say, ten times the stock's recent daily volatility and is reversed within one or two days, then compare the price with a second source or the day's traded high and low. Check prices against corporate-action calendars so a genuine split is not flagged as an error. And make every backtest report show its result without its best five days: when 97% of the return sits in two days, that one line tells you to look at the data before the strategy.