Algo & Quant traders

[SimpleMeLike]

Looks like there were significant drawdowns with your click-click system?

Hello ironchef

Yes, I experience losses and drawdown while clicking.

Nothing you can do about drawdown, but keep on clicking and get out of it.

There is no other choice buddy.

Whatever, the other choice is , I damn sure do not know about it. So I just stare at the chart and click click click click.

There is not other choice.

 

[globalarbtrader]

Good Morning p0box4,
So I think best to curve fit the entire data series so you have chance to make money every year.

View attachment 338989

No... you're wrong.

And you're right - not in a million years should you be allowed near an algo.

Curve fitting makes performance better in the past, but usually makes it worse in the future when you're actually trading.

Let me try and explain it like that. Imagine that you had a strategy that went long and short on different days ot the year. And then you curve fitted it to the performance of 2023; so that for example if on the 20th april 2023 the market went up, you would go long. And if on the 21st april it went down, you would go short.

The performance of that strategy would be the best - it would never lose money on any single day of the 2023 backtest.

Would you really expect that strategy to make money in the future? For that to happen the market would have to go up on the 20th april in 2024, down on the 21st and so on. The most likely outcome is that your up/down/up/down... fitted strategy will do about the same as random coin flipping. But because you have to pay trading costs, you would actually lose money.

That sounds crazy, but when you optimise, that is effectively what you are doing. Assuming the past will be exactly the same in the future. The more you optimise, the better the historic back test looks, the worse it will actually do in the future.

(Your other problem is that one year of data probably isn't enough to fit this strategy, but that's another post)

But you want to have the best possible performance in the future, and you wouldn't be satistifed with a backtest that wasn't overfitted. The issue here is you have massively unrealistic ideas of what sort of performance you can achieve. Making a profit every single month just isn't achievable by the vast majority of traders, even professionals, except for those in the high frequency trading space. Pretty much the best hedge fund on the planet, Rentech, loses money about two months in every year.

GAT

 

[Frederick Foresight]

Curve fitting makes performance better in the past, but usually makes it worse in the future when you're actually trading.

Just to put a finer point on it and for the sake of clarity, I don't think there is anything wrong with curve fitting per se. In fact, I think it's essential in developing a repeatable strategy. It is overfitting that leads to problems going forward.

As I see it, curve fitting is like manufacturing an off-the-rack suit to fit a general segment of the population reasonably well. However, overfitting is going bespoke; it will fit only one wearer particularly well (i.e., the look-back period) and virtually everyone else (going forward) not at all.

 

[SimpleMeLike]

No... you're wrong.

And you're right - not in a million years should you be allowed near an algo.

Curve fitting makes performance better in the past, but usually makes it worse in the future when you're actually trading.

Let me try and explain it like that. Imagine that you had a strategy that went long and short on different days ot the year. And then you curve fitted it to the performance of 2023; so that for example if on the 20th april 2023 the market went up, you would go long. And if on the 21st april it went down, you would go short.

The performance of that strategy would be the best - it would never lose money on any single day of the 2023 backtest.

Would you really expect that strategy to make money in the future? For that to happen the market would have to go up on the 20th april in 2024, down on the 21st and so on. The most likely outcome is that your up/down/up/down... fitted strategy will do about the same as random coin flipping. But because you have to pay trading costs, you would actually lose money.

That sounds crazy, but when you optimise, that is effectively what you are doing. Assuming the past will be exactly the same in the future. The more you optimise, the better the historic back test looks, the worse it will actually do in the future.

(Your other problem is that one year of data probably isn't enough to fit this strategy, but that's another post)

But you want to have the best possible performance in the future, and you wouldn't be satistifed with a backtest that wasn't overfitted. The issue here is you have massively unrealistic ideas of what sort of performance you can achieve. Making a profit every single month just isn't achievable by the vast majority of traders, even professionals, except for those in the high frequency trading space. Pretty much the best hedge fund on the planet, Rentech, loses money about two months in every year.

GAT

Hello globalarbtrader,

I understand what you mean. Thank you for explaining.

Why optimize at all if its such a risk of curve fitting? Just do not optimize any of the data series or 50% of the data series.

Here is algo back test I built in 2022. I run in 2023, and loss -$15,000 cash money in the ES.

I have no clue on earth why this algo failed on me. I thought I be rich in 2023 to scale up to big size ES contract, yet 2023 was painful as heck.

I optimize from 2006 to 2019, and the rest of years out of sample.

And still lose money.

Now what do I do?? What is the solution for this?

And look at this year so far, barely made $1000 full ES contract.

How can anyone make money trading algo when things like this may happen?

Now, keep in mind, I have check if this algo is on 250 days for 2023 every morning, knowing I am in drawdown and losing money!!!!! That is painful.

 

[SimpleMeLike]

Just to put a finer point on it and for the sake of clarity, I don't think there is anything wrong with curve fitting per se. In fact, I think it's essential in developing a repeatable strategy. It is overfitting that leads to problems going forward.

As I see it, curve fitting is like manufacturing an off-the-rack suit to fit a general segment of the population reasonably well. However, overfitting is going bespoke; it will fit only one wearer particularly well (i.e., the look-back period) and virtually everyone else (going forward) not at all.

Hello Frederick Foresight and @globalarbtrader, @wxytrader, @p0box4 and the rest.

Why did my algo I built in 2022 lose me -$15,000 cash in 2023?

I only optimize about 60% of the data set from 2006 to 2019, the rest was out of sample.

Please tell me what happen.

Thank you,


https://www.elitetrader.com/et/threads/algo-quant-traders.379241/page-21#post-5959048

 

[Frederick Foresight]

Hello Frederick Foresight and @globalarbtrader

Why did my algo I built in 2022 lose me -$15,000 cash in 2023?

I only optimize about 60% of the data set from 2006 to 2019, the rest was out of sample.

Please tell me what happen.

Thank you,


https://www.elitetrader.com/et/threads/algo-quant-traders.379241/page-21#post-5959048

Assuming you used the same strategy consistently throughout the losing period, that the loss was a significant percentage of your capital, and that there was a representative sample of trades during that period, then I would guess that you either overfit or underfit the data in your look-back period. And that you did not do sufficient forward testing to validate the method.

That's my guess, anyway.

 

[hilmy83]

Hello Frederick Foresight and @globalarbtrader, @wxytrader, @p0box4 and the rest.

Why did my algo I built in 2022 lose me -$15,000 cash in 2023?

I only optimize about 60% of the data set from 2006 to 2019, the rest was out of sample.

Please tell me what happen.

Thank you,


https://www.elitetrader.com/et/threads/algo-quant-traders.379241/page-21#post-5959048

I'm guess your strategy sucks to begin with dude.

It seems you make money in very volatile periods (2008, 2019-2020), then pretty much in comatose for other years..

Then the market got fed up with your nonsense and gave a "fuck your volatility" in 2023.

 

[SimpleMeLike]

I'm guess your strategy sucks to begin with dude.

It seems you make money in very volatile periods (2008, 2019-2020), then pretty much in comatose for other years..

Then the market got fed up with your nonsense and gave a "fuck your volatility" in 2023.

LMFAO HAHAHAHAHAHAHAHAHH that is FUNNYYYYYYYYYYY

Good laugh.

LOL, you right!! Not sure what happen in 2019 to 2022 in the market, but I can sure use that again.

 

[SimpleMeLike]

Assuming you used the same strategy consistently throughout the losing period, that the loss was a significant percentage of your capital, and that there was a representative sample of trades during that period, then I would guess that you either overfit or underfit the data in your look-back period. And that you did not do sufficient forward testing to validate the method.

That's my guess, anyway.

Hello Frederick Foresight,

I am not sure what happen either. It had good size of sample trades. I did not overfit.

I think @hilmy83 said it right. guess the ES market changed or something, but it made me lose -$15,000 in cash in 2023. DAMN, I thought it work too.

That is why I give props to anyone with an algo that makes money, cause better to make money then lose money.

 

[p0box4]

Hello Frederick Foresight and @globalarbtrader, @wxytrader, @p0box4 and the rest.

Why did my algo I built in 2022 lose me -$15,000 cash in 2023?

I only optimize about 60% of the data set from 2006 to 2019, the rest was out of sample.

Please tell me what happen.

Thank you,


https://www.elitetrader.com/et/threads/algo-quant-traders.379241/page-21#post-5959048

There is absolutely no way of knowing what went wrong without knowing more about the algo you've made. Instead of asking us, why don't you dive in the executions and data and figure out what went wrong? We don't have access to the data, you do, and I guarantee you the answer is in the data.