Looking further back than a year ?

Discussion in 'Strategy Development' started by futurecurrents, Feb 6, 2012.

  1. I'm talking about intra-day trading.....

    There's been a lot of talk recently about how the markets are changing. Day trading is dead, etc. Traders that did well one year often see those returns widely vary as the market changes. Eric P has testified to this as have others here.
    It seem to me that between the changes in volume, the dynamic nature of algorithms, the changes in market participants and always novel economic conditions, that worrying about how a strategy did two years ago is not really relevant My thinking is that yes, finding a strat that would work in those older conditions is a good idea, if only to have ready to use should market behavior change, but the most important is the more recent time-frame, especially when talking about day trading.

    I want to believe this as my strat backtests well last year has forward tested well now for two months, but not the year before, where the action and volume looked different.

    Thoughts?
     
  2. MGJ

    MGJ

    Why don't you trade it, as they say, in real life using real money at 1/10th position size, for 2000 trades? If it's a really spectacular system, this little test will give you some very comforting confirmation, and it only costs you 3 months of your life.

    On the other hand, if there is a weakness in your system or in your backtest simulations of your system, causing it to perform poorly, the cost to you has been reduced by a factor of ten!
     
  3. Thoughts …

    A. I believe there are many arguments for and against your approach... however, if you're going to rely just on the latest period, make sure it contains a statistically meaningful number of trades. How many are you talking about? (Again, there are different views about what represents a "statistically significant" number of trades... me? I like 300 min, 1000+ better )

    B. And recognize, from what you've written above, that the strat is highly likely to break down again at some point; so decide now how you'll detect this so that you can stop trading it before you've given back all your profit (assuming you end up having some, that is!). The historical test performance stats are useful here as a reference, and techniques like Monte Carlo (not universally accepted here on ET!) are useful too IMHO.

    C. It's also a good exercise to try to understand why it might work now but not previously? What aspect of the market is different, and why would this make a difference? Can you relate the periods when it works or not to a "signal" of some type?

    D. If you can't find a fairly recent historical period (apart from the last 12 months) where your strat does reasonably OK in an out-of-sample test, you're faced with the dilemma of not knowing any better whether your results are a chance curve fit, or whether they indeed take advantage of some new aspect of intraday market microstructure that's never existed before. If you're a HFT player with $millions invested in infrastructure, you could indeed be dealing with the latter; if on the other hand you're a retail guy/gal trading from home (or from a "click trading" prop firm), what's more likely? So follow the earlier suggestion and trade "1/10" of size until you have better certainty ...
     
  4. ammo

    ammo

    take a look at the range in spx for the 1st qrtr,year after year
     
  5. ammo

    ammo

    all the white vertical lines denote jan 1,so the bar on the white lines is the 1st qrter,the randomness of that small or large range may effect your system.you can do the same with the 2nd,3rd 4th and compare that with your systems range of profits/drawdowns
     
  6. Thanks for the thoughts guys.

    "Why don't you trade it, as they say, in real life using real money "

    I'm doing it. It's automated. So far so good, but although it's a day trading system it only generates a couple of trades a week. I can increase the trade number but the returns don't change much.

    So this ties in with

    "if you're going to rely just on the latest period, make sure it contains a statistically meaningful number of trades"


    The problem is that to generate a large number of trades I have to go back into times where the market action was different, negating any validity of looking back further, other than to prepare for a similar future period.

    "And recognize, from what you've written above, that the strat is highly likely to break down again at some point; so decide now how you'll detect this so that you can stop trading it before you've given back all your profit (assuming you end up having some, that is!). The historical test performance stats "

    Great advice. I must admit that at this early stage in my learning I've been looking for more ways to make profitable trades instead of determining a stop-trading the system criteria. One way I was thinking of was to put a trend line on the performance graph and stop trading when the trendline is broken. But it seems that even when the system doesn't work it doesn't lose much.

    "It's also a good exercise to try to understand why it might work now but not previously? What aspect of the market is different,"

    I'm trying to figure that out. I've looked at volatility, intra-day range and trend. But the big difference I see is volume. The system is on the Q's and the volume in 2011 was only 77% of the volume in 2010. And that volume was decreasing all year. In 2011 the volume leveled-off.

    "If you can't find a fairly recent historical period (apart from the last 12 months) where your strat does reasonably OK in an out-of-sample test, you're faced with the dilemma of not knowing any better whether your results are a chance curve fit, or whether they indeed take advantage of some new aspect of intraday market microstructure that's never existed before."

    I should do this. I must admit to being afraid to do so. But again, if the past is different, what good would it do? That being said, what I'm doing is nothing new.

    "take a look at the range in spx for the 1st qrtr,year after year"

    Thanks, I've looked at ranges and there does not seem to be a correlation.

    One thing on my to do list is to create an intraday trendiness indicator. Because although intraday range is known, I don't easily know what's going within that day. I'm using TS, I'm thinking that just a simple number of new five minute daily highs/lows, should give a good indication?
     
  7. It's all about risk.

    When you're trading with real money, you can always set a rule to change your position size to zero if the theoretical equity curve goes below a moving average.

    Trend following your equity curve.

    :D
     
  8. How do you deal with "whipsaw" around the equity curve (i.e. getting in just after winners but before losers, or out after losers but before winners)?