Machine designed strategies. Do they work?

Quote from amazingIndustry:


Lol, intraday and not simulating fills? Paying bid+ask all the time? And how would he model slippage? I thought I invested time but I feel I wasted it...;-(
More...

I agree, gmst you need to improve on that (but i dont like the ego over there). I simulate limit fills with price penetration. Market fills get filled at the worse part of the spread 50% of the time, if there is high volume, it's scaled up linear based on volume vs avg volume. My own software has a model for the spread for each contract ID which is constructed by looking at the spread live (DOM) or via tick data. Stop orders are filled similar to market orders except for higher losses in volume spikes.

Sorry i cant recommend software that does it, i just build my own.
 
braincell - Sometimes the spread dominates the effectiveness of a strategy.

If you want to be buying the bid, and selling the ask, you need to simulate fills without the price level being consumed entirely.

This is where it starts to get harder to build a backtester.
 
I feel I am repeating myself. So according to your reasoning you should indeed monitor butter prices in Bangladesh, they did a terrific job at predicting s&p index returns out of and in sample. Good luck with that.

My job is done, I came in to warn newbies to spend 90% of their time monitoring their screens and learning market dynamics rather than (over) fitting parameters to historical prices (whether automated or not). If Machine Learning works for you then all the power to you. My last question to you is, with all the fanciness and the ability to spit out 100s (it should actually be 1000s) of tradable strategies you should already be sitting on several dozens of working strategies. If that is the case, my respect to you. It just does not sound that way...maybe my hunch is wrong...who knows. Good luck everyone.

Quote from braincell:

What's the matter with everyone i talk to that has as much as experience as you do, which i respect, thinks that every backtesting app built ever doesnt include real-world market aspects like spreads, slippage, etc. My one does, and disappearing bids would be important if i was an HFT, not a 1 hour holding period. But ok anyway, what you're saying is that future cannot be foretold.

Ok now we get to walk forward testing. If we get two OOS samples, 25% of data each, we can find 100s of systems and see if positive OOS1 data correlates to positive OOS2 data and the scatter graph looks nice, what, you're saying that means nothing - again? My opinion is if there is correlation in OOS1-OOS2, we can safely assume that when we get only OOS1 that is positive, OOS2 (which now becomes live trading) should have a statistical chance of holding up which is equivalent to the previous OOS1-OOS2 tests. Anything wrong with that logic?



I agree, i did some testing a year ago and the fuzzyness was profound. I'd rather trade the VIX heh.
More...
 
Quote from braincell:

I agree, gmst you need to improve on that (but i dont like the ego over there). I simulate limit fills with price penetration. Market fills get filled at the worse part of the spread 50% of the time, if there is high volume, it's scaled up linear based on volume vs avg volume. My own software has a model for the spread for each contract ID which is constructed by looking at the spread live (DOM) or via tick data. Stop orders are filled similar to market orders except for higher losses in volume spikes.

Sorry i cant recommend software that does it, i just build my own.
More...

What you write is correct, I can always improve my trading by buying bid selling ask if get a good fill rate - earning the spread rather than paying on many trades is definitely going to improve my trading performance.

However, as an independent trader, I have finite time resource. Currently, I am spending most of my time on building strategies and understanding market phenomena that persist over time. At this point, I just want to go after low hanging fruits, so it calls for prioritization on my part. Limit order fill simulation is way way down my priority list currently, and I don't see it changing for the foreseeable future. I am convinced that it is possible to make consistently more than 10 million a year as an independent trader without worrying about limit order fill simulation. After that maybe yes, one needs to look into that. But I am way off from that point - so don't want to put my energy into something that is very hard and I don't need it as of now.

I respect that you developed your software but I hope you see my perspective here.
 
So now I did not get your posts? By the way, attention to proper accounting for spreads, changes in the order book, and slippage is equally essential in intra-day futures trading. Not sure what experience you have that you claim it is not. Yes I have an attitude, but I am not here to massage it, I get my feedback and worked hard to be rewarded by the market not by posting opinions and sharing experience. Sorry if it does not sit right with you.

As mentioned in the previous post, I came in to warn newbies to not lose focus over screen time by paying too much attention to touts that promise the holy grail selling software.

Job done, I am out of here. Sorry if I came across as getting personal...was not my intention. I just get very agitated when certain individuals harvest amateur's curiosity to market inappropriate products.

Quote from gmst:

Buddy you got an attitude I admit it! Please read my post I replied to braincell that I am talking in reference to futures trading. And mind you I am not trading for a fund wherein I need to buy/sell notional of 100 million within 2 minutes. I am running an independent trading account and for anything less than 10 million notional, my modelling of 1-2 tick slippage with spread is good enough.

Don't jump to conclusions mate. and calm down if you want a proper discussion. I don't mind if your are massaging a bit of your ego with your posts here - thats fine - but everything in moderation :)

Anyways, the point is you didn't fully get my post. I asked for a "somewhat not too bad" software that will simulate fills, not great top of the line product.
More...
 
Quote from Rationalize:

braincell - Sometimes the spread dominates the effectiveness of a strategy.

If you want to be buying the bid, and selling the ask, you need to simulate fills without the price level being consumed entirely.

This is where it starts to get harder to build a backtester.
More...

This is one of the primary reasons I just want to concentrate on strategies that have a high average trade profit. In ES anything less than 50$ a trade and I am not interested because currently I don't have the technology to ensure that I will be able to correctly simulate fills and slippage.

I am sure if someone has a good simulation fill software, even positive expectancy strategies with 10$ average trade profit for ES would make sense to run. Actually, as one goes towards less average trade profits, number of trades increases, your variance decreases and it results into a smoother uptrending equity curve - provided and it is important that your fill simulation is correct. And this is where real difficulty is.
 
apologies if I seem to get personal, was not my intention. I got carried away because I take it as one of my jobs to protect the unassuming from snake oil salesmen and fishy testimonials, at work as well. Nothing bugs me more than the BS that is played in this industry. But I stand by my opinions that are grounded in my extensive experience as quant trader. Hope it works out for you and others.

Quote from braincell:

I agree, gmst you need to improve on that (but i dont like the ego over there). I simulate limit fills with price penetration. Market fills get filled at the worse part of the spread 50% of the time, if there is high volume, it's scaled up linear based on volume vs avg volume. My own software has a model for the spread for each contract ID which is constructed by looking at the spread live (DOM) or via tick data. Stop orders are filled similar to market orders except for higher losses in volume spikes.

Sorry i cant recommend software that does it, i just build my own.
More...
 
Quote from amazingIndustry:

So now I did not get your posts? By the way, attention to proper accounting for spreads, changes in the order book, and slippage is equally essential in intra-day futures trading. Not sure what experience you have that you claim it is not. Yes I have an attitude, but I am not here to massage it, I get my feedback and worked hard to be rewarded by the market not by posting opinions and sharing experience. Sorry if it does not sit right with you.

As mentioned in the previous post, I came in to warn newbies to not lose focus over screen time by paying too much attention to touts that promise the holy grail selling software.

Job done, I am out of here. Sorry if I came across as getting personal...was not my intention. I just get very agitated when certain individuals harvest amateur's curiosity to market inappropriate products.
More...

Actually, I agree with most of what you are saying especially about warning newbies.

As per my experience about fills rate in futures - if you are trading spot FX on EBS, you can do 10 million on major pairs with zero slippage. Correct. An independent trader doesn't do more than 10 million a clip on fx. Also you can do 100 contracts on ES with zero slippage - most of the time of the day. Correct? An independent trader typically doesn't do more than that. This is where I am coming from, and I clearly wrote this in my last post. Again, I am sorry but you are not reading carefully, most likely in the heat of the moment.

Btw, I don't mind little bit of attitude if I am learning something. But I found it odd that you replied to me without carefully reading what I wrote. Cheers.
 
Quote from amazingIndustry:

So according to your reasoning you should indeed monitor butter prices in Bangladesh, they did a terrific job at predicting s&p index returns out of and in sample. Good luck with that.
More...

Nope, incorrect. I never mentioned two data series which have no logically apparent relationships (butter and S&P). OOS1 and OOS2 are a time series of the same instrument, same strategy, same conditions and their logical similarity is thus infered. Two very different things. If your point is that inference is impossible, i'd like an explanation for apparent structure in the scatter graphs describing the properties of OOS1 and OOS2, for example net profit which tends towards the upper right quadrant.

I might be sitting on some strategies but because of my skepticism which may not be apparent from my posts, i am still only paper trading them. I also need to improve my platform, and need more data to be sure about what i'm saying. Sampling problems are always a concern. I also hate making claims that later go under water.
 
but dont you see that this is in summary the whole fallacy of curve fitted models. You end up with statistically significant results, let's say the causal relationship is there, yet you are still not having enough confidence to put the strategy on. Why? I claim because you know well that the strategy's parameters were fitted to historical time series. You tell me, "well, the out of sample holds up equally well", but you still know in the back of your mind that the dynamics of the strategy were not derived from logical reasoning that you derived from experience about market dynamics but that it instead originated from a mathematical optimization tool. In the end you know very well that markets often enough utterly disrespect mathematical predictions that originated from optimization tools. A perfect analogy is the option writer who does reasonably well until the one moment when hell breaks lose. He knows the day comes but not when. Yet, he engages in the strategy even he knows his life span is finite and bankruptcy is inevitable, instead relying on probabilities. This is the precise reason most bank traders writing options get fired every now and then. The only one single reason to engage in this business is an unethical one: The free option (get lucky to survive couple bonus seasons before fuxxing up). Someone else must clean up after them and they knock on another door for more of the same scam. Independent traders do not have such luxury.

If you feel comfortable knowing that your strategy dynamics break down sooner rather than later then curve fitting to a degree is not evil, but you must most likely know that those with the highest confidence in their models are the ones that derived the dynamics from actual trading and risk taking experience and not mathematical mode. Heck, I say this with an advanced degree in Stochastic Calc and Financial Engineering. But I deeply believe in what I just said.



Quote from braincell:

Nope, incorrect. I never mentioned two data series which have no logically apparent relationships (butter and S&P). OOS1 and OOS2 are a time series of the same instrument, same strategy, same conditions and their logical similarity is thus infered. Two very different things. If your point is that inference is impossible, i'd like an explanation for apparent structure in the scatter graphs describing the properties of OOS1 and OOS2, for example net profit which tends towards the upper right quadrant.

I might be sitting on some strategies but because of my skepticism which may not be apparent from my posts, i am still only paper trading them. I also need to improve my platform, and need more data to be sure about what i'm saying. Sampling problems are always a concern. I also hate making claims that later go under water.
More...
 
You must log in or register to reply here.
Back
Top