The best trading strategies of PRC for Prorealtime – Backtests update & Rankings
- This topic has 63 replies, 13 voices, and was last updated 2 months ago by
.
-
5 days seems like a reasonable time to close a zombie trade…
96 days will be zombie madness… 🙂
1 user thanked author for this post.
Hi PRC traders,
Hello Meta Signals Pro,
first of all, congratulation for launching this tool, it is something that a lot of us we’re looking for. In general, I spend a lot of time doing backtestings of all the bots i can, and therefore measuring the performance, so that i can do some kind of “asset allocation” between all the bots i work with. This tool would save me a lot of time, but unfortunately i must give a negative feedback (by now).
I’ve registered, paid the feed and tested the first published results, the top ones. As it is announced, the backtestings have the maximum time depth (this means 10years, 15 years…), and probably this is a problem. Just to examples:
1.Most profitable (and free source) of the analized algos of https://best-trading-algos.com/ is pic01.
As you can see, you get a x10 profit running it since 2010.
But what about if you run it just last year, with current market conditions (pic02)?
Ops, you get a bankrupt. Perhaps if we valuate between doing backtesting of 15y vs 1y, something changes.2. Second example, the second most profitable bot on https://best-trading-algos.com/ is “End Of Day – YEN M15” (pic03)
Doing an 8 years backtesting, we get some impressive returns.
What about if we run it only last 12 months (pic04)?
Results don’t seem to be interesting. Even we detect a high drawdown last months.
I love the long term concept in investments. But in home-made trading we cannot hold an strategy that is lossing 6 months in a row, 12 months, etc etc, even if the backtesting indicates that is a good option in the long term.
What i mean, and all of us already know:
- Past performance does not guarantee future returns. Having backtestings of latest 15 years may guide to error. It is more trustable getting a 20% performance on 2022, rather than getting 40% performance on 2010.
- While setting a new strategy on production, you don’t check the last 20 years performance. The mandatory condition to get in real mode an strategy if that it had a good behaviour with last 12 months data.
It isn´t useful doing backtestings with data longer than ¿3 years? (5 years? ___FILL HERE YOUR OPINION____). Market evolves and the probability of getting 15 years’ ago conditions is minimum in comparision with last 3 months market data.
As we have checked in these two examples (the top 2 on the ranking), doing a backtesting with 10 or 15 years is very dangerous (figures are there). It is more accurate doing a backtesting with only last months data.
What I suggest:
https://best-trading-algos.com/ is a very good idea that i engourage to improve. But i proposse to add a view: a part from having the current backtesting with the maximum historical data (10/15 years), to include also a ranking based on the backtesting of the last 6 months. This would be more realistic and would help to save time to developers, investors, gessing out exactly and quicker with algorithm is best suited to current the market conditions.
Thank you.
3 users thanked author for this post.
Sorry, i got a typo on the cell number, my bad.
1 user thanked author for this post.
Hello,
I think what @nicolas, @ThaNoizy and I (@alfredocuaresma) suggest is that, although the idea of evaluating all the possible algos is quite (QUITE) interesting, it should be executed in another way. In the current one, it is mostly useless (unless you have a Delorean to travel to the past).
The problem
best-trading-algos is getting a powerful ranking of algos based on the latest 10 years backtesting.
Imagine you get a first algo with this performance: +150% in 2013, and -5% next years till 2023 = +105% performance.
Imagine you get a second algo with this performance: 0% in 2013, and +5% next years till 2023. = +45% performance.Following the ranking of best-trading-algos, you would focus directly on the first algo. And taking into consideration that in the last 9 years, it has lost 5% every year , it makes sense that this year you will lose another 5%.
Otherwise, if you choose the second algo, as it has generated each one of the last 9 years a +10% performance, you can consider that you have good possibilities to get a 10% this year.
As a registered user, I have deeply studied the ranking of best-trading-algos and I consider this is its problem, it puts on the top of the ranking the algos with better returns on the last 10 years, and most of them follow the behaviour of the first example.All of us, when doing a backtesting, we check that in the last 12 months (maybe 6 months or 3M, depending on trader preferences) we won’t lose money when getting the algo in production. We don´t care if that algo earnt a 3.000% 10 years ago, what we want firstly is not to lose our money in the next 3 months. This is key.
Market data of latest 6 months are more reliable than the latest 60 months. Paradoxical? Not at all. Markets are continuously evolving, its conditions changing like volatility or actor´s behaviours, and other factors which are quite different every year. Doing a backtesting of the last 10 years means giving the same value to market conditions of 2013 than market conditions of 2023. That is probably the key error of best-trading-algos (but fortunately there is a solution).
Every trader has a feeling of how many months he should use when doing a backtesting, and we´re talking a range between 3 and 18 months. More than 18 months we would probably be including incrementally noise that would generate wrong results.
The solution
Solution for best-trading-algos is easy: to repeat the backtestings with less market data. After this, in the best-trading-algos web, to add 4 new columns: “performance only in the latests 3M”, “perf 6M”, “perf 12M”, so we can order the ranking with these 3 columns. Also adding a 4th column, the backtesting date (i.e. 15/Oct/2023) , it will help so we can filter algos or old backtestings.
We already know all algos can generate a 3.000% if doing a “””right””” overfitting, so the current information on the web is useless. First thing an investor wants when looking for an algo is just to know how much money has that algo earnt/lost in the short term, last year. Having these metrics in your platform, me and many people would join best-trading-algos it thankfully.
I got some other extra ideas to best-trading-algos, like adding a 5th column named “drawdown 12months”, or an easy Sharpe Ration based on return and drawdown. That would make best-trading-algos a great solution for many users of ProRealTime on all platforms on the world.
Thanks.
3 users thanked author for this post.
Thank You for your well written, easy readable / understandable post with good ideas .
I am even more ‘short term’ than you highlight above … I’d love to (also) see a performance column over 1 month / Last 4 Weeks.
Many Thanks also for your hard work thus far; this Topic would reach a whole new level if you might be able to do as described in the post above?
2 users thanked author for this post.
Hi Alfredo,First I would like to thank you sincerely for contributing to this topic; it has been such a painful work;
And pushing us to build a better tool;Please see my answers to your relevant questions below;
ChrisHello,
I think what @nicolas, @ThaNoizy and I (@alfredocuaresma) suggest is that, although the idea of evaluating all the possible algos is quite (QUITE) interesting, it should be executed in another way. In the current one, it is mostly useless (unless you have a Delorean to travel to the past).
To me, generally speaking in trading, statistics on past datas is the only tool we have to ground a strategy.
More precisely in algotrading, all the tools that we have (WinRate, Drawdowns…) are also always based on the past.
I am currently studying some prejective tools but they will always used past data anyway.
The problem
best-trading-algos is getting a powerful ranking of algos based on the latest 10 years backtesting.
As for I read here and there, backtest of less than 200KU were not “relevant” and on the contrary pretty much misleading;
This is why I standardised this to 200 000 time units.
Please note that is the TF is in minute the time span is not 10 years anymore 😉Imagine you get a first algo with this performance: +150% in 2013, and -5% next years till 2023 = +105% performance.
Imagine you get a second algo with this performance: 0% in 2013, and +5% next years till 2023. = +45% performance.
Following the ranking of best-trading-algos, you would focus directly on the first algo. And taking into consideration that in the last 9 years, it has lost 5% every year , it makes sense that this year you will lose another 5%.Otherwise, if you choose the second algo, as it has generated each one of the last 9 years a +10% performance, you can consider that you have good possibilities to get a 10% this year.
As a registered user, I have deeply studied the ranking of best-trading-algos and I consider this is its problem, it puts on the top of the ranking the algos with better returns on the last 10 years, and most of them follow the behaviour of the first example.
All of us, when doing a backtesting, we check that in the last 12 months (maybe 6 months or 3M, depending on trader preferences) we won’t lose money when getting the algo in production. We don´t care if that algo earnt a 3.000% 10 years ago, what we want firstly is not to lose our money in the next 3 months. This is key.
This corresponds exactly to my algotrading experience; launching paid or free algos and then getting an immediate drawdown that is painfully despairing (specialy if you have bought the algo for a couple of months…)
The problem is that you cannot know of course when the next drawdown or the next underperformance will happen.In your example the fact that the strategy has performed lately gives you no prediction of how it will REALLY behave in the coming 3 months right?
However the angle of the equity curve gives you hints though; if it quite regular with a constant slope it is a good start.Personnaly I have stopped looking for a universal and eternal strat; it does not exist because the market is not modelisable.
However, I noticed that strategies does not randomly become underperforming; this is probably because the market does not change chaotically from a certain market condition to another;
I became pragmatic:
1.sorting the strats that are too erratic from the others
2. and apply my StratSENTINEL on the most profitable ones. So that I can see when they get off-road.
StratSENTINEL is incorporated on the tool (every 3-4 months) for the most interesting strats paid and free.
The great solution would be to be able to have a shorter updates I reckon.
Market data of latest 6 months are more reliable than the latest 60 months. Paradoxical? Not at all. Markets are continuously evolving, its conditions changing like volatility or actor´s behaviours, and other factors which are quite different every year. Doing a backtesting of the last 10 years means giving the same value to market conditions of 2013 than market conditions of 2023. That is probably the key error of best-trading-algos (but fortunately there is a solution).
Yes I agree on this totally;
Every trader has a feeling of how many months he should use when doing a backtesting, and we´re talking a range between 3 and 18 months. More than 18 months we would probably be including incrementally noise that would generate wrong results.
The solution
Solution for best-trading-algos is easy: to repeat the backtestings with less market data. After this, in the best-trading-algos web, to add 4 new columns: “performance only in the latests 3M”, “perf 6M”, “perf 12M”, so we can order the ranking with these 3 columns.
I will try to add an additionnal images of shorter period; but it is a new development for the tool;
Also adding a 4th column, the backtesting date (i.e. 15/Oct/2023) , it will help so we can filter algos or old backtestings.
Interesting but this is just huuuuuuge manual work;-) I’ll see what I can do.
We already know all algos can generate a 3.000% if doing a “””right””” overfitting, so the current information on the web is useless. First thing an investor wants when looking for an algo is just to know how much money has that algo earnt/lost in the short term, last year. Having these metrics in your platform, me and many people would join best-trading-algos it thankfully.
I got some other extra ideas to best-trading-algos, like adding a 5th column named “drawdown 12months”, or an easy Sharpe Ration based on return and drawdown. That would make best-trading-algos a great solution for many users of ProRealTime on all platforms on the world.
Thanks.
2 users thanked author for this post.
Find exclusive trading pro-tools on 
