Possible explanation for this?

Discussion in 'Strategy Development' started by logic_man, May 17, 2012.

  1. I trade the ES and while I don't trade it 24 hours/day, I do track what would have happened if I did trade every instance of my set-up, with the rationale that if I ever went fully automated, I'd like to know what to expect outside of the hours I do trade, which are 6AM to 6PM ET. Also, if a trade started in the off hours and is "in progress" once 6AM rolls around, I wait for that trade to complete before taking another trade in that same direction, to avoid taking a "trade within a trade", which can happen.

    Anyway, over the past year, I've had about 100 trades that occurred during the hours I trade and 100 outside of those hours. That's apples-to-apples trades as much as possible, i.e. they exhibit approximately the same characteristics relative to the parameters in my model.

    In 97 trades during the core hours, the average trade was a gain of 4.97 points, stdev of 8.55.

    In 92 trades during the off hours, the average trade was a loss of 0.33 points, stdev of 7.09.

    I've known about this for a while and I've known to just stop trying to get into trades past 6PM, but I just can't understand why the exact same algorithm would be so successful for 12 hours and then a net loser for the next 12 hours. Statistics would say there is basically no probability that these two sets of trades have the same average outcome, so what gives?

    My best guess is that the traders from whom I take those ~5 points per trade are just not in the market overnight. This makes sense to me because the day session has many more retail traders than the overnight session and I would assume that those traders are the net losers in the market. Even that leads to the unanswered question of why I am able to take those points from those traders when they do show up and why those points aren't taken from those traders by the same traders who would take that 0.33 points from me in the overnight session. It's all just a bit odd, at least to me.
  2. how important of a component to your trade signals is volume and/or momentum? the tape is much easier to paint through the dark hours. have you looked over those 200 trades to see where the night session "went wrong" compared to what you expect to see from your RTH trades? have you tried to see how the night session trades would have done with a looser stop in place, or with tighter targets?

    maybe your algo gets sucked in on the night volume action "looking good" in the between US close and UK open (or at the uk pre/open), only to get whipsawed back to break even on the low night volume and tendency to sweep the stops, play the range, or throw out head fake reversals, *just* before any real move happens going into the UK session proper and next day.

    i've often wondered if people (algos) specialize in the off hours exclusively, someone must, as it's when the funk, magic, and shake 'n bake happens that the 8AM traders are left to wake up and deal with. there's no denying there are moves there that could be traded, nearly every night.

    interesting numbers though, comparing the two sessions, hope you can get to the bottom of it and let us know what you find. maybe you are right, it's a much smaller pool at night, with the same large sharks in the dark, more likelihood to find yourself lower on the food chain than a quick suckerfish in the regular trading hours.

  3. I'd filter out all trade and quote data from 4:15pm est to 6pm. Rerun and see if your results are normalized.
  4. Volume is irrelevant to me, but I've got an algorithm which defines momentum and it's the same in both sessions. I don't want to mess with the rules to try to make the overnight session profitable, but I just found it very curious to see the persistent discrepancy. I do think that overnight is even more dominated by institutions than the day session, although institutions do the overwhelming amount of volume in both.

    Most of the overnight trades originate between 2AM and 4AM ET. Those trades average nearly a 2 point loss per trade. The trades from 6PM to 2AM are actually an average winner of a little over 1 point, but by 6PM I'd rather disengage from the markets than look for that one point, at least now while I am not automated. So, it's really between 2AM and 4AM that the rules don't work. Then, they start to work again after 6AM. In fact, trades entered between 6AM and 7AM are, on average, 11.75 point winners, which is by far the best of any 1 hour of the day. This makes sense because I think the origins of most "trend days" occur during that hour and the big boys position themselves then to distribute to retail traders later. "Early bird catches the worm", I suppose. The only bad thing is that there have only been 6 of them in the last 11 months since I've been keeping time of day stats at this level of detail.

    What adds another level of curiousness to the situation is that these same rules appear to work just fine on the Euro and oil from 3AM onward. I am in the early phases of rolling out the strategy to those markets, so I am doing more paper-trading than real trading in them, but it's already noticeable that this is happening. Of course, it could change as the sample size grows. My thinking is that those two markets are front and center of European traders' attention during those hours, so my algorithm can exploit the same behavior in those markets as it does when the US wakes up and the ES becomes the center of attention.
  5. I looked at that, but the number of trades which rely on data from 4:15PM to 6PM is quite small, so the overall results don't really change.

    I know that time period tends to be very slow, but is there another reason why you thought it would be important?
  6. JoCracket


    Where can I go to find out what you guys are talking a bout?