forex-scalping.com

Strategy testers rely on historical data. Many platforms download that history from the broker's own server, and there is no guarantee that demo and live servers store the same data.

In some cases the demo server runs a separate feed, with a different set of liquidity providers or a different pricing engine. Spreads on the demo feed are typically cleaner. Gaps in the data may be filled in differently. Weekend quotes can appear on one and not the other.

A backtest using demo history can therefore show results that the live account can't reproduce. The difference might be small on average and large during specific periods such as news or low-liquidity hours.

A cautious approach is to compare the history from both servers for the same period. Export one-minute data for a week from the live and demo accounts and check high and low values bar by bar. Where they differ, the question is why.

Third-party tick data from providers such as Dukascopy or similar is often used for testing instead, and it has its own quirks, mainly that it comes from a different liquidity pool from the broker's.

None of the sources is the true market. Each is a sample.

Which data source has the closest match to live fills in testing?

I got caught by this when I tested a breakout system on my broker's demo history and then switched to live. The demo data was noticeably cleaner around the Asian session, with fewer of the little spikes that live had, and some Sunday evening bars on the live server simply didn't exist on demo at all. The live results were worse within the first couple of weeks, and the gap came almost entirely from those quiet hours.

When I did the comparison you describe, exporting a week of one minute bars from both servers, around one bar in eight had a different high or low, usually by a fraction of a pip but sometimes by two or three during news.

What has matched my fills best is recording my own tick data from the live account with a small script and testing on that. It took a few months to build up a useful history, but it's the only data I trust now. I still use downloaded third party history for longer tests, but I treat those results as a rough first pass.