my EA bleeds during news, should it just stop trading around red folder events

[Nathaniel]

every backtest looks great once you remove all the times it would have lost. add a news filter, then a friday filter, then a monday filter, and eventually youve curve fit a strategy that only trades the three hours it happened to win historically.

[carterw]

the troll names a real danger, filter creep is how people overfit by exclusion. the discipline is: only add a filter for a mechanism you can explain, not just a losing period you noticed. 'spread blows out during news so my entries are bad' is a mechanism. 'it loses on tuesdays' usually isnt. that distinction keeps the filter honest.

[itsliam]

whats a reliable calendar feed to pull from, ive heard some of the free ones are wrong or late?

[silvercloud]

use a well-known economic calendar with a documented data feed rather than scraping a random site. and crucially, test that your filter actually blocks trades around a known past event in the strategy tester before trusting it live. a filter that silently doesnt fire is worse than no filter because you think youre protected.

[oceanbreeze]

the 'verify it actually fires' point is underrated. ive seen news filters that looked configured but the event-time parsing was off by the broker server offset, so it paused at the wrong moment entirely. always confirm against a real historical event.

[thomas_x]

for what its worth, after adding a clean news filter my EAs equity curve got noticeably smoother without hurting total return much. the news trades were nearly a wash on average but added huge variance. cutting variance for similar return is a good trade in itself.

[cloudyvision]

smoother equity curve with similar return is also exactly what overfitting produces in-sample. forward test it on events it has never seen before you celebrate.

[thomas_x]

fair, and easy to settle: keep the filter, run the EA forward on demo through the next month of news events, and compare filtered live behaviour to what the unfiltered version would have done. if the filter still helps on unseen events, it was a mechanism, not a fit. let the forward data decide.

[lowkeysam]

filter for a reason, then forward test it. clear.

[neonpilot]

this is one of the most useful threads ive read here in a while the 'mechanism vs noticed losing period' line is going straight in my journal.