Could it be this easy?
- Thread starter farmerjohn1324
- Start date
That's a typical remark that you often read. But it makes no sense to me.
In 10 years markets can/do change. So these results might not be representative for the actual market anymore. What worked then might not work anymore today.
I would study the behavior of the system in various different market conditions. The way the strategy adapts itself and gives good result in any market condition is more important as you never know in advance which type of market you will trade in future.
I checked an intraday strategy on about 1,000 trades (period was roughly 1 year long) and the results are still inline with real trades 10 years later.
So the expectancy of the strategy is still comparable with the checked test. Only the profits rose as a result of some adaptations done as I understood better and better the relation between the market and the strategy.
Algo trading works perfect. But the only problem is that you have to create that algo.
And what is the issue with checking 100 trades? Too much work for many people.
It does not take that long, and as you do it:
1) create a catalog on the type of shenanigans and keep count. For instance: Suspect fills, data anomaly, after the fact fills etc.
2) keep track on what type of movements happened that day. Trending or consolidation. You will quickly get an idea how the Algo performs on those type of day.
3) look at the top 10 winners and losers. See the details on how they worked.
4) look at the drawdown length, and why it happened, i.e. why so many in sequence.
5) run it for a 12 month period, outside the one you did. and again different 12 month period. Then look at the worse 12 month period and the best.
6) run it on different instruments. Repeat the above.
So it is not that simple as seeing some interesting back test result and trading it live. But it is pretty straightforward to investigate if you are willing to do the work.
"I would study the behavior of the system in various different market conditions"
Best way to see how a system behaves in various different market conditions is to do a 10 year back test. Actually the more years you have the better, so 10 years is a minimum.
This is applies to a general trading system that is not dependant on recent market behaviour. If you have developed something to exploit a recent pattern, then the performance in previous years is less relevant.
Last edited:
I agree with this. This is like using one type of clothing for year around weather for 10 year or 10 locations. Or one club for all course, in all weather, all locations on the green.
And what people do is "over fit" based on too much generality, so you get something that either:*
a) does really good in some and not really bad in others. (day traders).
or
b) does so-so in all markets, but reasonably positive. (e.g. spread traders)
or
c) a lot of really good and really bad which you reject because you are just wasting time and comms.
Just like a good trader trades the market that is presented to them, takes a profit or loss, based on what the market presents, so should an Algo, usage be based on what the market presents.
So the trick is to write the basic Algos, (plural), like a 4 iron, to use the golf metaphor, then a Wedge, and a putter. Then write the Meta Algo code that know which one to use based on the market. Or use the on going performance to adjust, turn off or on the various "clubs".
A lot of work in any case, but doable with persistence and a non flat learning curve. So, again, "no", it is not that simple just finding a single Algo and testing the heck out of it.
*Perhaps the desire for an "All Season" algo, is based on the false notion that you set it and go on vacation for 3 months or a year. In reality, there is constant adjustments and monitoring, or "farming" as they call it.
Last edited:
For a daytrader that's pure nonsense.
One simple example: I started daytrading with a 2 points stop ES more than 2 decades ago. Because the ES went up x-fold that stop should now be 5 points to represent the same risk reward ratio. The margins now are also much higher then when I started daytrading because the notional value of the ES went up hugely.
I developped my strategy more than 2 decades ago on 1,000 trades representing roughly one year of data. That strategy still works today, so more than 2 decades later. That's hard proof for me that you don't need 10 years.
So the trick is to write the basic Algos, (plural), like a 4 iron, to use the golf metaphor, then a Wedge, and a putter. Then write the Meta Algo code that know which one to use based on the market. Or use the on going performance to adjust, turn off or on the various "clubs".
A lot of work in any case, but doable with persistence and a non flat learning curve. So, again, "no", it is not that simple just finding a single Algo and testing the heck out of it.
Exactly.
I don't trade on an MA, rsi, fibonacci, support/resistance or stochastic.
I recognize patterns, and for each pattern the "selfmade indicators" have different rules. So the rules for an indicator in pattern A are different from the rules for that same indicator in another pattern. That's the algo ( rather one of...).
If the a pattern switches to another pattern things start from zero again, rules for each indicator can change again. The swich of pattern is the Meta Algo.
I call it more a modular system with modular subsystems.
I agree with this. This is like using one type of clothing for year around weather for 10 year or 10 locations. Or one club for all course, in all weather, all locations on the green.
And what people do is "over fit" based on too much generality, so you get something that either:*
a) does really good in some and not really bad in others. (day traders).
or
b) does so-so in all markets, but reasonably positive. (e.g. spread traders)
or
c) a lot of really good and really bad which you reject because you are just wasting time and comms.
Just like a good trader trades the market that is presented to them, takes a profit or loss, based on what the market presents, so should an Algo, usage be based on what the market presents.
So the trick is to write the basic Algos, (plural), like a 4 iron, to use the golf metaphor, then a Wedge, and a putter. Then write the Meta Algo code that know which one to use based on the market. Or use the on going performance to adjust, turn off or on the various "clubs".
A lot of work in any case, but doable with persistence and a non flat learning curve. So, again, "no", it is not that simple just finding a single Algo and testing the heck out of it.
*Perhaps the desire for an "All Season" algo, is based on the false notion that you set it and go on vacation for 3 months or a year. In reality, there is constant adjustments and monitoring, or "farming" as they call it.
@virtusa
It sounds like my Step #1 is to truly learn how to backtest properly. Which it sounds like means manually going back through the chart for let's just say 10 years.
I need to find out if my platform and/or computer will even support that much tick data. For some reason, I could backtest on 1-tick for one week and it took only a few seconds. But I tried for one year and it never gave results and NinjaTrader closed.
Try live forward testing.