I wrote a bot and gave it 25k to trade
- Thread starter _terminus_
- Start date
Do you know how it performed in 2008-2009 and in 2011? For most long strategies, 2008-9 are killer years. But you seem to use some trend filters so you might just avoid these periods.
The data I have at the moment doesn't go that far back, so I can't answer this question. However, you guys are getting me curious so I will get more data and do the necessary backtests.
I do remember those were very bad years, so testing them might be worthy, although I'm not sure about the meaningfulness of market conditions more than a decade ago.
I do remember those were very bad years, so testing them might be worthy, although I'm not sure about the meaningfulness of market conditions more than a decade ago.
I'd recommend it. It's not too far into the past, market was not dramatically different from now regulation-wise. I think we're likely to see similar patterns reappear in the future.
Spurred on by some ET members, notably @SPYAlgoTrader, @d08 and @traider, I decided to conduct a further back in time test for my main strategy. So I bought 1-min historical data for NQ from 2007 to date and did some experiment.
By using my usual approach of plotting the % return of trades, trying to get price independent results, a problem immediately appeared around 2007-2008: a 170% drawdown that supposedly would have wiped me out (picture below).
So I started to change the strategy parameters and conditions, trying to smooth out those nasty peaks and valleys leftmost on the curve, to no avail. Whatever change I tried, the results only got worse. On the one hand this heartened me, because it means the strategy is already good enough. But certainly not enough to leave it at that.
Then, the enlightenment.
During all this time I failed to realize a simple fact: summing up the percentage returns, without taking price into account, biases the results because the NQ price has changed substantially over time. On April 2007 it was priced 1843. Today it's worth 11326. So in other words, a 1% return in 2007 was very different from 1% in 2020. More than 6 times different.
So I tried another approach. Instead of % returns I plotted the $ returns, scaling all them up to current NQ price. So for example if a trade returned $100 in 2007, I multiplied those $100 by (11326/1843) = $615. In this way, a max drawdown of $9900 would have incurred into, over 13 years.
In other words, the test answers the question: What would have happened if the price of NQ had remained fixed at the same value as today, and I used this strategy (which can open max 5 positions, 1 MNQ contract each) repeatedly? I consider the answer quite reassuring. Yes, there is a long flat period from 2007 to 2012 that barely produced any profit. But at least the max DD is not that disastrous.
Results in the graph below.
One interesting fact that can be noted is that the $ curve looks more parabolic relative to the % one. This could logically be expected if position size increased over time, which is not the case: always 5 positions, always 1 contract. And price scaled up as if it remained constant.
For this reason, it looks like the efficiency of the strategy is increasing over time. This could possibly be explained by the ever rising importance of tech stocks. Just a hypothesis off the top of my head.
In conclusion, the test gave me ideas and suggestions to further improve the strategy, which I'll try to do before long, and I can only thank the members that prompted me to perform it.
Please feel free to criticize and spot any flaw in my reasoning.
By using my usual approach of plotting the % return of trades, trying to get price independent results, a problem immediately appeared around 2007-2008: a 170% drawdown that supposedly would have wiped me out (picture below).
So I started to change the strategy parameters and conditions, trying to smooth out those nasty peaks and valleys leftmost on the curve, to no avail. Whatever change I tried, the results only got worse. On the one hand this heartened me, because it means the strategy is already good enough. But certainly not enough to leave it at that.
Then, the enlightenment.
During all this time I failed to realize a simple fact: summing up the percentage returns, without taking price into account, biases the results because the NQ price has changed substantially over time. On April 2007 it was priced 1843. Today it's worth 11326. So in other words, a 1% return in 2007 was very different from 1% in 2020. More than 6 times different.
So I tried another approach. Instead of % returns I plotted the $ returns, scaling all them up to current NQ price. So for example if a trade returned $100 in 2007, I multiplied those $100 by (11326/1843) = $615. In this way, a max drawdown of $9900 would have incurred into, over 13 years.
In other words, the test answers the question: What would have happened if the price of NQ had remained fixed at the same value as today, and I used this strategy (which can open max 5 positions, 1 MNQ contract each) repeatedly? I consider the answer quite reassuring. Yes, there is a long flat period from 2007 to 2012 that barely produced any profit. But at least the max DD is not that disastrous.
Results in the graph below.
One interesting fact that can be noted is that the $ curve looks more parabolic relative to the % one. This could logically be expected if position size increased over time, which is not the case: always 5 positions, always 1 contract. And price scaled up as if it remained constant.
For this reason, it looks like the efficiency of the strategy is increasing over time. This could possibly be explained by the ever rising importance of tech stocks. Just a hypothesis off the top of my head.
In conclusion, the test gave me ideas and suggestions to further improve the strategy, which I'll try to do before long, and I can only thank the members that prompted me to perform it.
Please feel free to criticize and spot any flaw in my reasoning.
I'm no expert, but I honestly don't think that data from 13 years ago should be used to evaluate a strategy today. Buy low and sell high will always work, but even in my few years of trading, I see things happening more and more that didn't years ago. There are many more head fake moves. (ie. price is pushed down to sweep the entire book of 10 levels lets say in a micro second), and this doesn't appear to me to have happened years ago as much. Since you bought minute data, I do believe these types of moves would be visible even on 1 minute bars.
My point is that the algos are programmed much more differently now than 13 years ago, and more predatory in my opinion. In that same time frame, spoofing has gone from being a real thing, where a "thick" level actually meant something, to now it being obvious that whatever size is quoted in the DOM needs to be taken with a grain of salt.
So I'm not sure how affected your strategy is by these tricks, but if you're using minute data, I'd say the game has most certainly changed over the years.
With reference to this, I find it interesting and wonder what this says about your strategy. In this time frame, there certainly was enough of the same type of action we have now. There was some good direction, with the 2007-2009 down move, and then it was a bull market, and some deep retracements along the way. Of course 100 points meant so much more when the value of the ES was only 1200, vs. what 100 points means now.
But I still wonder how this time period didn't make many money. Its like some traders say they can't make money if the ES is stuck in a 10 point range all day and require at least 20 or 30 point moves.
So I wonder if perhaps your strategy isn't in some way flawed to take into account the range of moves in that time period. (When I say flawed, I mean programmed to take into account 200 point days for NQ vs. 40 point days) Today, you wouldn't think of using a 5 point NQ stop with a daily range of 200-300 points, but years ago it was quite acceptable. If the inputs in your strategy don't account for the average range, then I could see how it appears that for years it didn't make money, but could easily have given a simple adjustment to account for volatility, or simply the average daily range. Its pretty much the same thing you point out when talking about dollar value vs. percentages.
I just fail to understand how for years it didn't make money and all of a sudden took off after the 500th trade.
My point is that the algos are programmed much more differently now than 13 years ago, and more predatory in my opinion. In that same time frame, spoofing has gone from being a real thing, where a "thick" level actually meant something, to now it being obvious that whatever size is quoted in the DOM needs to be taken with a grain of salt.
So I'm not sure how affected your strategy is by these tricks, but if you're using minute data, I'd say the game has most certainly changed over the years.
With reference to this, I find it interesting and wonder what this says about your strategy. In this time frame, there certainly was enough of the same type of action we have now. There was some good direction, with the 2007-2009 down move, and then it was a bull market, and some deep retracements along the way. Of course 100 points meant so much more when the value of the ES was only 1200, vs. what 100 points means now.
But I still wonder how this time period didn't make many money. Its like some traders say they can't make money if the ES is stuck in a 10 point range all day and require at least 20 or 30 point moves.
So I wonder if perhaps your strategy isn't in some way flawed to take into account the range of moves in that time period. (When I say flawed, I mean programmed to take into account 200 point days for NQ vs. 40 point days) Today, you wouldn't think of using a 5 point NQ stop with a daily range of 200-300 points, but years ago it was quite acceptable. If the inputs in your strategy don't account for the average range, then I could see how it appears that for years it didn't make money, but could easily have given a simple adjustment to account for volatility, or simply the average daily range. Its pretty much the same thing you point out when talking about dollar value vs. percentages.
I just fail to understand how for years it didn't make money and all of a sudden took off after the 500th trade.
As for the number of years required to develop a strategy, I don't have a clear answer either. I suppose some strategies are built to benefit from knowledge of a great deal of past data, whilst other (i.e. some kind of machine learning) are tuned to work only on very recent price action.
I initially developed this strategy using data from 2013 to 2020, and sure enough it wouldn't have worked very well before that period. But at least it wouldn't have suffered disastrous losses. This is more important to me. Now, with more data I can decide to make adjustments.
It's rather the norm, not surprising, that a strategy behaves very differently outside the time period used to develop it. So the question remains: how far back should you go? I guess there will never be a definitive answer to that, because trading = uncertainty.
This strategy uses 5 min bar data.
I initially developed this strategy using data from 2013 to 2020, and sure enough it wouldn't have worked very well before that period. But at least it wouldn't have suffered disastrous losses. This is more important to me. Now, with more data I can decide to make adjustments.
It's rather the norm, not surprising, that a strategy behaves very differently outside the time period used to develop it. So the question remains: how far back should you go? I guess there will never be a definitive answer to that, because trading = uncertainty.
This strategy uses 5 min bar data.
What do you mean, "easier"?