/biz/ - Business & Finance

>>9600163
How many of your models are actually profitable? Did you create your algo trading strats by yourself?

>>9600163
Favorite indicators?

>>9600262
i dont think you understand what quants are

What is the point of having 120 different models? Do you manually switch between them depending on the situation or is it automatic?

>>9600163
What was your alpha in the run up to Dec 17? How many hours a day do you devote to which element of trading?

Personally, I just hodled until Jan so was even against the market. 3 hours a day on the risk management aspects of mah model, and 90 mins relearning the math I flunked hard back in school

>>9600240
the ones running live average out on top over time. Obviously enough I gradually decommission those that go into prolonged draw dawn. There are multiple ways of doing this. For example, look into equity curve trading. Different things work for different people though. What I do personally is model out-of-sample drawdown (via monte carlo for example) and compare it to actual drawdown. If the actual drawdown exceeds the 95th percentile I decommission the system. You basically just have to keep coming up with new shit all the time. And yes, all of my stuff is mostly coded from scratch but varies greatly in complexity.

I'm also trading as well, although less of a professional and more of a hobbyist. Too much mathematical modeling tends to be more work than it's worth(at least for my small account).

>>9600163

So what type of models you using? LSTM? CNN? I've yet to find anything that out performs a simple mean regression strategy desu. Give the computational cost and the issues of supervised training, it needs to have a pretty strong performance boost over ARIMA or whatever.

I'm considering use of an A3C unsupervised learning setup to simply pick the parameters to the mean regression strat or take some manual actions, so for eg in a big dump it can react in a more intelligent way (like dump fucking everything). Seems like a good way to be able to more easily integrate fundamentals and sentiment too.

>>9600356
*trading algos

>>9600282
well quants aren't categorically against TA, although it is frowned upon in most places. You can easily take a pure TA system and make it profitable by filtering out false-positives/false-negatives with another model on top if the market allows. The main problem with TA is complete lack of objectivity by most participants

>>9600286

Aggregation. You're lucky if a specific model is hitting +2% or so over flipping a coin, so you run a bunch with different seeds and aggregate the results. Brings it way up. Much more expensive to run and train of course.

>>9600419
You mean primary model? Generally speaking, if we are talking about weak data (of course if you have access to very strong predictors I would use some kind of ensamble/boosting method and be done with that) I would never use neural nets or similar setups as the primary model.

The easiest way to use machine learning in finance is in a meat learner setting. For example, suppose you have a something generating a buy/sell signal already and now you want to improve this, put a machine learning model on top. Or another common example is setting position sizes via machine learning. And so on, you get the idea.

My background is in abstract maths/explanatory maths so I generally tend to use a ton of pure numerical methods, decompositions, signals processing, etc. And then improve those with machine learning on top

I am too brainlet for this

>>9600457
I don't do this btw. Risk of ruin would be immense. The main reason behind this many models is strategy diversification. Taking pure data mining based stuff aside, say you really like mean reversion. Mean reversion is usually characterized by high win rates, stable but generally speaking tiny profits and really fat returns tails. Meaning that when it loses it loses big. One way to avoid this (arguably, which I generally don't agree with) is some stop-loss setup. Another way to avoid this is reducing variance by just trading uncorrelated strategies/assets.

>>9600526
* meat learner -> meta learner lol. Fucking insomnia

>>9600419
why not use PSO or something similar for parameter optimization? Although, I would really advice against any of this, you are doing glorified curve fitting essentially. If you end up doing this make sure you run a ton of robustness tests and do your backtesting correctly

>>9600424

Yeah, nah, not really. Using mean regression based strategies (common for quants) is certainly common, and similar to someone using a Bollinger bands, RSI or EMA based strategy. Pretty much anything to do with candle patterns or meme lines (especially that Elliott waves bs) is out. It just doesn't work - I've tried a bunch of different approaches personally, it's a dunning kruger IMO. There's no real improvement in trying to take all the usual TA memes into account.

Not to say that TA is all bad for human traders. It might be a bit of tea leaf reading, but the human brain is still a great pattern recognition engine so some pseudoscience props can maybe help it along. With sufficient practice (ie staring at enough charts) you can get good at picking trades based on intuition. I just doubt it's really the TA itself that's providing the answers.

The real secret is in risk management and capital protection. If your rewards (averaged) exceeds your risk, and no trade is going to wipe you out you don't actually need to be placing perfect trades. Everyone focuses on prediction too much IMO.

>>9600651
as I said, it has to be quantifiable somehow (hence the name). One simple way of looking at it is asking yourself "Can you systematize it?", If you can, go ahead, backtest it (properly, not like 95% of literature out there). Obviously candle sticks don't work. That doesn't mean you can't exploit a weak entropic system via some indicators. Also doesn't mean that those indicators will work in every regime, all of the time, etc. Its all about risk/reward as you say

>>9600646

Yeah, like I say just playing with it we'll see.
Having it do meat learning (lol) seems like a good idea.

More DSP based stuff is on my list actually, I want to get into wavelets, see if using signal decomposition yields anything interesting.

Honestly though, I think the REAL issue is we're all conceptually off base with all this. It's not actually a function being iterated over, ie each state x isn't just a product of states x-1....x-n. So really, it needs to be approached more from a game playing angle, an imperfect information game with unknown players of which the price at current state x is the result of their competition. Hence why I think attempting to tie in SA a lot more is the direction I'm going.

Back testing is tough though, sentiment data in sufficient volume is expensive and kinda shoddy. Might end up just setting up a couple of scripts to scrape various sources for a year or so. Partly why I want to try A3C is for the unsupervised learning aspect of it.

>>9600725

Yeah 100%. If you can't quantify it, you can't systemise it.

>>9600163
What language do you code in?

Any advice for someone who has also studied maths and stats looking to get into this?

>>9600163
what do you think about enigma catalyst? Is there a better platform for a newbie quant to get started?

What risk management tools have you guys implemented? Using a pretty simple trigger based on trend but feel like I could do more. Did save me from a 10% dip this week tho

>>9600811
Wavelets are good for detecting structural breaks (summing eigenvalues) and stuff like that (for mean reversion again). Similar decompositions approaches will work too, like QR or singular level decomposition. Just don't use any of it for signals smoothing or real time stuff. You'll run into endpoint anomalies.

I would personally say signals processing is everything. Guys from rentech agree lol. Theres a reason for simons buying out ibm's speech recognition department. If you are able to smooth the signal in real time you can do eveything else, including forecast it as well. Its def non-trivial though. My most profitable algos are all variations of signals approaches and various decomposition functions.

>>9600847
Combination of matlab/python/R, sometimes cpp for production. I'm a shit programmer/engineer, my brain doesn't really function in this way too well. So if you are worried about the programming part, don't be. I would also say unless you are planning on doing HFT related shit the language is really not all that important.

Re starting out, go here

> https://quantocracy.com/books/

pick out a book and get your hands dirty. If you have a maths/stats background shit will be a LOT easier than for someone starting out from scratch. Start experimenting with a bunch of shit and I'd say half a year in you'll be good enough, it gets easier from there

>>9600847
I would especially recommend Chan's books for beginners (skip his first book though unless you are a complete noob)

>>9600651
This. I've looked at 1000s of charts. and the best indicators are PSAR for determining trend reversal, MACD histograms and RSI for picking entry, and BBands to determine moves up or down. Combining that with price history, and status of indicators at different intervals, and a volume moving average makes it very difficult to loose.

The thing is the indicators are becoming offset now, b/c too many normans are using them.

>>9600546
I second.
Still interesting to read English words that make no sense.

>>9600163

If you're looking for a job at an extremely well paying firm, email me cryptoassist@gmail.com. They pay really really well did I mention that.

>>9600163
Nice overfit backtesting performance

>>9601129
wouldnt "normans" using your same trading strategy actually help you? i don't understand.

>>9601674
heh
>>9601678
> overfitting a fir filter...
> I can use big words too mom
but nice try brainlet

>>9600163
This thread is depressing, I wrote what I guess is something like an evolutionary algo that spitballs candlestick TA and keeps breeding and mutating the best indicators together based on backtest performance. I've had it running on Binance with play money since January and it's averaged .7% daily. But, I have zero cognizance of, like, linear regression, risk modeling, monte carlo sims, or any of the other big words you used in this thread. Obviously my bot could just be running on luck and could fuck itself to zero tomorrow, I have no way of knowing. I do at least pass the buy signals through several virgin chunks of historical data at the end so I'm hopefully not just overfitting myself to death.

I thought I was clever when I wrote it but I'm pretty sure I'm a caveman with a ray gun who thinks he's hot shit but doesn't actually know what the fuck he's doing.

>>9601810
.7% is really good

The best investors keep it simple

>>9601810
Wouldn't that have like tripled your money? Seems pretty good.

What language do you use and why?

>>9601810
A few suggestions. How does your equity curve look like, is it mostly smooth? You can calculate Spearman's rank correlation coeff, something above .90 is solid. Look into your e-ratio too. Try some robustness tests. For example, try altering some of the parameters slightly, does your equity curve change significantly? Try altering exists/entires slightly. Is your algo affected by seasonality? Derive various statistics like p/e, correlation coefficients, e-ratio, drawdown, etc. for both in-sample and out-of-sample data. Do they correlate? Try running your strat on a different correlated asset (if you work with btc, try running it on eth, etc). Do results differ dramatically? There are other tests you can perform but those are fairly easy to setup and should get you started

>>9601711
The entire 'TA is a self fulfilling prophecy' meme is perpetuated by idiots that don't actually trade.

There's a limited amount of liquidity for taking a trade at a certain price, if everyone gets the same signal it's a race to whoever gets their order filled first. This becomes a substantially bigger issue for algo traders / hft traders because the difference in getting a trade is seconds or fractions of a second.

>>9600163
OP I’m a researcher in deep learning. I do mainly natural language processing and information retrieval. What kind of models do you use? Where can I get started with this? I find it pretty interesting.

>>9600262
Indicators != Statistics

>>9602066
Nice. Do you do research into deeplearning or do you use deeplearning in your nlp research? And I use anything and everything available. From pure modelling (exploiting inefficiencies like momentum), data mining, anomaly detection, information theory based approach (like measuring system entropy and picking decision points based on that), pure numerical methods, etc. Generally its really about being creative and coming up with new stuff.
Pick up a good book have a look here >>9601015 and go from there

Yo, OP. Aspiring quant here and have been working with another guy for six months.
Would love to talk more if you have the time. Do you have a throwaway email so we could talk more?

Btw Gaussian filters are bomb

Two separate questions assuming you are in US. How did you handle paying taxes on every transaction? Are you registered as a trader?

>>9600594
I'm really interested in this stuff but haven't done much research yet. I'm correcting that mistake right now actually, you made me google meat learner though haha

>>9602177
2 Things. I picked up Marcos de Prados book. Quite insightful. I mainly algo trade using DSP and other mathematical techniques to do regression analysis. I believe de Prados' point on backtesting is extremely true. I believe algo traders can seriously benefit from backtesting on generated charts based on the composition of sinewaves from charts. for example, decompose a timeframe and get means and Standard errors of the 5 most dominant cycles you can extract, either through MESA, SSA, EMD, groetzel, whatever, and then use a simple algorithm to build a brownian candle chart, with the 5 waves changing in amplitude randomly based on the mean and SE, and backtest on it

2. Lets linkup, get >>9601904 in a discord, and scheme.

>>9602204
sure drop me a line: biznessanon[at]protonmail.ch please don't send dick pics though if at all possible, thanks
>>9602217
Nope Europe, Running swiss registered fund. Pay income taxes in country of residence (also europe) on money I "pay out" to myself. Everything else is almost tax free.
>>9602240
Look into Lopez de Prado's new book, he talks a lot about "meat" learners lol

>>9601711
yes and no. If everyone is thinking the same it is favorable but then you have bots that watch everyone and throw everyone off by placing sell/buy orders at the right prices. Since the indicators are based off price the bots skew it. Some calibration of the indicators can fix this.

>>9601735
Forgot to take my daily dose of self-loath. OP how much you make?

>>9600163
hey man you've made some interesting points here regarding your strategies around algotrading/price models. not totally related but in terms of project, do you take Chainlink seriously?

What's your annual return?
If it's less than 100%, why aren't you buying and holding?

>>9602272
100%. Especially his point re: information leakage. The thing is all of this stuff is pretty well known in the industry, Aronson talked touched upon many of Prado's points years ago. And regarding backtesting, I would argue it would make more sense to rather analyze results via some approach, but synthetic data makes a lot of sense too. Drop me a line, I hate discords and all of that, too much of a time waster, my email is above
>>9602308
a good amount, lets just say I had it made already way before btc became a think worth thinking about

>>9600163
Speak to me about NapoleonX (NPX) The leader of the project does what you did for BNS Paribas. Any opinions on project lead Stephane Ifrah or do founder Arnaud Dortois?

>>9602406
* Bayesian approach. Good thing I'm no novelist...
>>9602382
You are assuming alpha alone. I do smart beta models as well, look into PMPT for example or stuff like passive-aggresive mean reverting portfolios. So I am buying and holding just with constant rebalancing. This tends to outperform just holding in most settings

Can u please build a robot trader fund? I would love to throw down on something like that.

>>9602406
Good for you :) Any thoughts on the future of crypto? Seems like volume and hype are both dying, probably for the better.

>>9602461
>>9602455
>>9602455

Stinky quantbag answer my question!

>>9602455
>>9602488
alright, alright calm down sugartits. I'm the wrong person to ask about various shitcoins, no idea honestly. I trade and hold based on financial data alone. No idea about those dudes you've mentioned either. However, one thing I can assure you of is that BNP's quants are solid, at least the ones I worked with.
>>9602485
Only thing I can say is this. Financial innovation is hard and financial innovation tends to get hijacked. Financial innovators focus too much on the mechanics and on singular features. They tend to miss the bigger picture. Financial innovators talk a lot about new paradigms. From recent memory, reminds me of the innovations that came with non-agency residential mortgage-backed securities and Gaussian Copulas. We all know how that ended. That said, from a speculative perspective, I don't believe we are anywhere near the top of the bubble.

>>9602285
Thanks for the suggestion, I'll pick it up soon

I really want to get into algotrading; Im already comfortable with maths, microeconomic theory, programming and stats (except for time series)

Im thinking about reading Empirical Market Microstructures by Hasbrouck and, after that, Asset Price Dynamics, Volatility, and Prediction by Taylor

Am I in the correct path? Is there anything else I need to know before i get to these books? Any other suggestion? Thanks for creating this thread, OP, ive been looking for some guidance for a while

>>9600651
I've seen studies where TA patterns backtest profitably