Dave
- #25
- Edited 5:40am Jun 20, 2023 1:56am | Edited 5:40am
The search for the Level 2 SL/TP is finally complete after almost 2 days. That took a painfully long time :-(
Recap: I let the MT5 optimizer not only try several SL/TP combinations, but let it try these combinations with 100 random buy/sell sequences. Then I calculated the median profit of each setting to filter out the overly lucky/unlucky results.
Good news! This worked like a charm. Look how smooth the calculated profit curves are:
The best setting appears to be L2_TakeProfitPoints=400, L2_StopLossPoints=1100. The corresponding profit distribution profile looks like this:
Since we only did 100 random runs per setting, the histogram looks a bit coarse, but it is still good enough to see that we already made a big improvement by just adding a single scale-in. We are now almost in profit, even just at Level 2!
I will set my poor old overheating laptop to work again, and now start searching for the best Level 3 settings. See you guys in few days...
Recap: I let the MT5 optimizer not only try several SL/TP combinations, but let it try these combinations with 100 random buy/sell sequences. Then I calculated the median profit of each setting to filter out the overly lucky/unlucky results.
Good news! This worked like a charm. Look how smooth the calculated profit curves are:
The best setting appears to be L2_TakeProfitPoints=400, L2_StopLossPoints=1100. The corresponding profit distribution profile looks like this:
Since we only did 100 random runs per setting, the histogram looks a bit coarse, but it is still good enough to see that we already made a big improvement by just adding a single scale-in. We are now almost in profit, even just at Level 2!
I will set my poor old overheating laptop to work again, and now start searching for the best Level 3 settings. See you guys in few days...
- #26
- Edited 9:10am Jun 20, 2023 7:52am | Edited 9:10am
If we would run the EA live with the settings found in post #25, what profit do we get? The graph above is a histogram showing the distribution of the profit of 100 backtests each with a different random buy/sell sequence. While -50 is the most likely result, we could be lucky and get +500, or it is one of those days and we go home with -500.
Luckily, you don't need to gamble. There is way to make sure we always get the average case. To do that we just run thousands of EA's simultaneously. Some will be lucky, others will not, but the outcome will be the average.
Of course, running thousands of EAs simultaneously is not practical. Therefore, I modified the EA to run all these thousands of strategies simultaneously internally. You can download version 1.1 from post #1. Each day we open both a buy and sell trade and apply the SL/TP rules to each of them. At times, we have 13 trades open!
Since hedging is wasteful (you pay commission and spread twice to be effectively still out of the market... why would you do that?!), I added a "netting" backend that adds up all the underlying trades and only makes buys/sells when it needs to update the net position. For clarity: this is build into the EA; you don't need a netting account at your broker.
To activate the behavior described above, set RandomSeed=0. If you set RandomSeed > 0, you get the behavior like before: one random buy/sell sequence. Let's try this with the settings we found so far:
This runs ALL random sequences simultaneously, giving us the average case:
As you can see from the deposit load graph, the EA is indeed running several strategies simultaneously. The backtest report tells us the profit is -100, which looks about right given the histogram in post #25 (note that the left leg is a bit higher/heavier, so while the mean is about -50, the average is not): 
Luckily, you don't need to gamble. There is way to make sure we always get the average case. To do that we just run thousands of EA's simultaneously. Some will be lucky, others will not, but the outcome will be the average.
Of course, running thousands of EAs simultaneously is not practical. Therefore, I modified the EA to run all these thousands of strategies simultaneously internally. You can download version 1.1 from post #1. Each day we open both a buy and sell trade and apply the SL/TP rules to each of them. At times, we have 13 trades open!
Since hedging is wasteful (you pay commission and spread twice to be effectively still out of the market... why would you do that?!), I added a "netting" backend that adds up all the underlying trades and only makes buys/sells when it needs to update the net position. For clarity: this is build into the EA; you don't need a netting account at your broker.
To activate the behavior described above, set RandomSeed=0. If you set RandomSeed > 0, you get the behavior like before: one random buy/sell sequence. Let's try this with the settings we found so far:
This runs ALL random sequences simultaneously, giving us the average case:
- #27
- Edited Jun 21, 2023 12:12am Jun 20, 2023 10:46am | Edited Jun 21, 2023 12:12am
WOW... I just had a great idea! What if instead of backtesting each setting with 100 different random sequences and then calculating the median profit, we apply the new version of the EA that runs ALL random strategies simultaneously? That would be 100x faster and still give me the average profit of each setting!!
So I aborted the L3 search, installed version 1.1 and tried again, this time using RandomSeed=0 (meaning: run all random sequences at the same time). That gave me the optimal L3 settings super fast: L3_TakeProfitPoints=500, L3_StopLossPoints=1200:
Running 1000 backtests with the settings above, gives the following profit distribution:
From the equity curve we already knew we the mean would be positive, but unlike what the mathematicians always say about averaging and grids, there appears to be no long-tail?! The distribution is slightly skewed to the negative side, but not as asymmetric as some want us to believe.
Can we really make money from a random strategy?!
So I aborted the L3 search, installed version 1.1 and tried again, this time using RandomSeed=0 (meaning: run all random sequences at the same time). That gave me the optimal L3 settings super fast: L3_TakeProfitPoints=500, L3_StopLossPoints=1200:
Running 1000 backtests with the settings above, gives the following profit distribution:
From the equity curve we already knew we the mean would be positive, but unlike what the mathematicians always say about averaging and grids, there appears to be no long-tail?! The distribution is slightly skewed to the negative side, but not as asymmetric as some want us to believe.
Can we really make money from a random strategy?!
- #29
- Edited 2:00pm Jun 21, 2023 1:30pm | Edited 2:00pm
Now that I found a way to backtest 100x faster, I decided to not only add a Level 4, but also revisit the previously found settings as these all somewhat affect each other. Since we now have a rising equity line, I switched from optimizing "Balance max" to optimizing "Sharpe Ratio" to aim for a more linear equity line rather than a "bumpy race to the highest equity".
These are the settings I found:
This gives this equity curve:
With these stats (Profit factor 1.38, Sharpe Ratio 0.76):
Considering this is a strategy that has no edge (it just opens both a buy and sell each day), the result is amazing!
The profit distribution profile is very clearly to the right of the zero-point. The long-tails the mathematicians warned us about are nowhere to be seen?! Did we just show that the mathematicians are wrong and Averaging Down actually works?!
What do you think?
Also, have some of you reproduced these or similar results? Perhaps not with the exact same settings or the AUDCAD, but following along the approach described in this thread, were you able to also find a setting that has a positive expectancy?
These are the settings I found:
This gives this equity curve:
With these stats (Profit factor 1.38, Sharpe Ratio 0.76):
Considering this is a strategy that has no edge (it just opens both a buy and sell each day), the result is amazing!
The profit distribution profile is very clearly to the right of the zero-point. The long-tails the mathematicians warned us about are nowhere to be seen?! Did we just show that the mathematicians are wrong and Averaging Down actually works?!
What do you think?
Also, have some of you reproduced these or similar results? Perhaps not with the exact same settings or the AUDCAD, but following along the approach described in this thread, were you able to also find a setting that has a positive expectancy?
Level 2 - SL/TP Now we are adding a second level. Once the first trade hit the stop loss, we now leave open the first trade and add a second trade in the same direction as the first trade. For both trades we now set new stop loss and take profit levels relative to the entry price of the second trade.
Hey Yoriz,
I looked closer at the TP/SL for each level. According to your 5th post, whenever a new trade is opened, the TP/SL for each previously opened trade is reset to the TP/SL used for the new level based off the entry price of the new level. If this is the case, when at the 4th level and the 4th trade hits it's TP (80 pips), how can the trades at Levels 1, 2, and 3 be in profit? At the time that the Level 4 trade is opened, the Level 1 trade is already down by 370 pips, the Level 2 trade is down by 360 pips, and the Level 3 trade is down by 200 pips. If, according to the 5th post, the TPs of the other three levels are now set at 80 pips based off the entry price of the 4th trade, the 80 pip TP exit for ALL trades won't cover the losses of the first three trades.
To verify this was not just a fluke, I tried again using a symbol that is less ranging and typically not used for grid trading: EURUSD.
Guess what:
This confirms we can really push the profit distribution profile to the left and get a positive expectancy!!
Can someone else please verify these results on their own system with their own broker?
Guess what:
This confirms we can really push the profit distribution profile to the left and get a positive expectancy!!
Can someone else please verify these results on their own system with their own broker?
