I Stress Tested a 1700% Strategy on CoinQuant AI [Full Backtest]

追加: 2026-05-31

[00:00:00]Most of the best looking trading strategies will fail the moment you put them live on the market. Even when the backtest was showing perfect positive returns, high chances the strategy might totally fail when deployed for live trading. Our last video's strategy revealed an amazing 1,700% in returns with commissions, spread, and total fees. Today, we are going to stress test it using the Monte Carlo method to see if this edge will survive through different market conditions. So, the backtest was repeated thousands of times randomly changing the strategy parameters just to test if random market changes will break our strategy. And I think this test should be done for every single strategy before trying it even on a paper account. Most of these promising backtests might just be a lucky statistical occurrence. Even when a strategy produces an incredible backtest, smooth equity curve, high win rate, impressive returns, it can still completely collapse the moment it goes live. And I think we all painfully know what I'm painfully referring to. Why?

[00:00:59]Because sometimes the equity curve is not skill, it's just luck and overfitting. And this is exactly where Monte Carlo robustness testing becomes extremely important. It's basically a statistical stress test for your trading strategy. Instead of trusting one single historical equity curve, we generate thousands of randomized simulations based on your strategy's behavior and ask much more important questions. What happens if trades arrive in different order? What happens if market conditions change slightly or even if we miss some trades. What happens if the price data is slightly noisy or you can have an extra slippage or some unknown conditions. Does the strategy still survive these conditions? So, the Monte Carlo testing tries to break this illusion of a fake winning strategy by answering whether the performance is genuinely repeatable. There are many ways to introduce randomness into a backtest to properly stress test a strategy. I'll show you a few of the most important methods before moving to an example. There are more techniques than these, but covering all of them would make this video too long, and maybe some of us will get bored at some point. The first test I like to start with is trade order shuffling. This one is uh one of the simplest but most revealing tests. We take all historical trades and randomly reorder them. Now, the total profit remains exactly the same. Same trades, same win rate, same profit factor, but the order changes, and this changes the drawdown structure dramatically. Maybe in the original backtest, the winning trades came early, and this is why the backtest was more positive. Now, imagine all the losses happened first. Suddenly, the strategy may experience a massive drawdown before recovering, and it might look more of a scary backtest. This test helps detect strategies that depend heavily on a lucky trade sequence. The second Monte Carlo test is trade resampling using a bootstrap method. This method, we randomly sample trades with replacement, meaning some trades may appear multiple times. Some trades may disappear completely. This creates thousands of synthetic equity curves, so we can check worst-case drawdown, confidence intervals, and probability of future underperformance. This is like instead of asking, "What happened historically?"

[00:03:18]we ask, "What could realistically happen next or in the future?" Another simple and useful method is random trade removal or skip trade test. In real trading, you will miss trades, maybe because of slippage, maybe uh execution delays, spread spikes, server issues.

[00:03:33]So, we randomly remove maybe 5 to 20% of trades from every simulation. Usually, a fragile strategy will collapse immediately when a few important trades disappear. A robust strategy should still survive, which means the average of the trades are good or is on the positive side, which makes this method a very powerful reality check. The next one is actually one of my favorites for detecting curve fitting. It's price path randomization. We slightly randomize the open, high, low, and closing prices. So, we randomize the data. For example, adding random noise of maybe plus or minus 0.1%, plus or minus 0.5%. Then, we rerun the entire backtest. If tiny changes in price completely destroy the strategy, that's a dangerous sign. It means the system is hypersensitive to exact candle structures or exact indicator crossings. And that is usually a symptom of over-optimization, also known as overfitting. In reality, markets are noisy, and the strategy should survive that noise. There are other ways using the Monte Carlo method like parameter perturbation, where we randomly change, for example, the RSI length instead of 14 between 12 and 16, for instance, or the EMA, for example, instead of taking one length of 50, we take a random sampled value between 45 and 55. All of this to test one idea. If changes happen randomly, so realistically, will my strategy survive and keep being profitable? And this concept is extremely important in systematic trading. As a general rule of thumb, we want around 1,000 simulations, preferably somewhere around 5,000 to 10,000 simulations to stabilize the percentiles and distributions. Also, never focus only on median performance.

[00:05:19]The important number is mostly the downside tail, the worst-case scenario, for example, the fifth percentile drawdown, because this represents a more realistic worst-case scenario. We're not saying this is the ultimate testing method, but usually a strategy that survives Monte Carlo stress testing is far more trustworthy than a strategy with a beautiful but fragile equity curve. Now, I will show you how I applied Monte Carlo robustness testing to our last tested strategy using CoinQuant, an artificial intelligence platform for backtesting strategies. You can find a link for the platform in the description of the video. It's a referral link for a free account, so you don't have to pay anything, but just to test the platform if it's something you'd be interested in. Okay, now we have access to the trades, the list of trades of the strategy. You can see it here. So, we have 15 pages of these trades. I can ask the AI here to check the trades of the strategy and estimate, for example, using Monte Carlo shuffling the results and the robustness. Let's send this message and see. I can see the message stress testing. It means that it's processing in the background, but I'm going to wait for the result to see what exactly happened. And just as a reminder, the equity curve, which is the green one, is smoother than the buy and hold. So, this is what we have experimented and backtested in the previous video. So, this is why we were interested, and I got a lot of comments regarding the strategy. Now, we're going to stress test it. I'm going to show you how it works using two or three Monte Carlo methods. We're going to see if it will keep this edge alive. So, the Monte Carlo testing is done. It took a bit of time because of the compute in the background. So, we repeated 10,000 shuffles, so 10,000 simulations in the background. And we discovered something.

[00:07:08]So, this is the overall feedback here in this card. So, the total cumulative profit and loss is still the same. So, this was expected. But, now we we revealed I mean, the backtest revealed this Monte Carlo shuffling revealed that the drawdown can be really scary. The probability of having a drawdown above 50% is around 62.5%, and the robustness confidence is now the score is around 55%. That's not high.

[00:07:36]You want something above 80%. So, we also have um, a lot of details here, everything that's that has happened in in the backtesting part. But, what I'm interested in now for this video is this. So, I have the metrics for the original backtest, then the Monte Carlo mean, the Monte Carlo median, and the Monte Carlo five the first percentile or the 5% and the 95%. And you can see here in the last column if the parameter we are measuring here, which is for example the final equity is path dependent.

[00:08:11]Let's check it out. So, the final equity is still the same, so it's not affected by the shuffling of the trades. Then we have the total return percentage. It's also the same as well, so these two are correlated, so it's kind of expected.

[00:08:28]The maximum drawdown, however, is very has very high variance when we use the Monte Carlo shuffling method. So, the original one was 19%, which was kind of acceptable. Although it's a bit high, but the Monte Carlo mean is around 80% drawdown, and the median is around 61%.

[00:08:49]And then we have 23% in the best case scenarios and 200% in the worst case scenarios. And if you recap the beginning of the video, this is what we should be focusing on. So, basically you are losing your account twice in the worst case scenario. This is already a parameter that will show you that this strategy is not ready to go live on the market. Then you have the sharp ratio, also the variations of the sharp ratio.

[00:09:15]Then we have the maximum loss streak.

[00:09:18]We have five and 11. Then we have [snorts] the profit factor. So, you can see that it's already a good reveal when you do those 10,000 simulations. It's showing you that you are still running a risky strategy. And this graph is really interesting here. It shows the maximum drawdown distribution across the 10,000 shuffles.

[00:09:36]We have very little safe cases. Only 22 simulations show the drawdown between 10 and 15% and the the maximum actually the highest frequency was for a drawdown of 60%. If you are uncomfortable with the histogram, we have the explanation here.

[00:09:51]AI is making it very easy for us. And then you have the sharp ratio and whatnot. So, all the other parameters.

[00:09:58]But, I don't think we need to go further in the backtesting. I think it's already enough to show us that the strategy is running with a high-risk set of parameters. Now, let's try another of the methods just for the sake of showing you another example. Let's try skipping some of the trades. What if some of the trades didn't go through the server correctly and so on. What it would make. Of course, you can combine shuffling and skipping and whatnot. But, for now let's keep it simple. I'm going to write in the chat box check the rates of the strategy and apply a Monte Carlo robustness test skipping some of the trades of the trades.

[00:10:39]Somewhere between 5% to 20% randomly using over 10,000 simulations, let's say. The skip robustness test is done. Took a bit of time, few minutes.

[00:10:54]So, the overall feedback is that the strategy is fragile. Strategy edge is outlier-driven, not statistically robust. And the robustness confidence is just 22%. What a disappointment. I thought we found a gem here. So, it run with 10,000 simulations randomly skipping 5 to 20% of 195 pair trades.

[00:11:19]So, this is the overall. And you can see again the let's check the total results here. So, we have the mean of the net returns is minus 6.95%.

[00:11:30]The maximum drawdown, the mean of the maximum drawdown percentage is minus 20% 21%. That's not very bad. It's not It's very close actually to what we've got, which is minus 19% uh in the original one. The win rate is around 40% in mean and the profit factor here. So, um this is the best 5% and let's check the worst. The worst is a net returns of minus 18% maximum drawdown of minus 28% win rate down to 38%.

[00:12:02]And not a very low uh not a very high profit factor. And of course, you have all the distributions, the final returns across simulations. As you can see here, minus 7.9%. So, this is the mode, which is the highest frequent case in here.

[00:12:16]So, I think it's showing how negative this is and only in rare cases we are on the positive side when we try to skip some of the trades. So, in other words, this strategy is relying on some of the trades that were making it profitable.

[00:12:32]When you remove those big positive trades, the strategy is mostly negative.

[00:12:37]So, um yeah, this is uh this is it. Uh I don't think it's uh as safe as we thought in the previous video. Okay, I will stop here for this video. As we saw, it was very smooth and complete to apply the Monte Carlo robustness testing using the CoinQuant platform. I would invite you to try it out. Check the link in the uh description. There's a referral link. Again, you don't have to go for a paid subscription. It's a free account. Try it out and if you have any feedback, let me know in the comments section. The CoinQuant team will be more than happy to check your comments and answer to your questions and even to your requests if you have any ideas you would like to add. Thank you for being patient and staying that long. Until our next one, don't forget to trade safe using the Monte Carlo robustness testing and see you next time.