AI - Artificial Intelligence

[Liberty]

AI algorithms actually are just an application of technical analysis, but in a diiferent context.

'History can predict future events'; in tech speak, make the machine calculate all possible correlations in a data set - & it WILL find some that are 'somewhat' predictive (middling R-square values). As it applies these correlations, we call it 'learning'. Of course, the 'machine' is only as 'smart' as the R-square of the correlation, and it's stability in an out-of-sample application; introduce it to a market-discontinuity, and it goes beserk :)

 

One of the theoretical arguments around HFT is that if your holding period is very small (nano-seconds), almost all your price gain will be attributable to market drift; and we can calculate the amount of that drift, using the Brownian Motion equations. Applied to AI, the more you can apply the Brownian Motion equations to an AI algorithm, the more accurate and stable it becomes. 

 

All things coming out of the 'investment' silo, and making the jump into other places.

 

SD

 

???

[Cigarbutt]

An interesting thread. One thing I notice is that the AI algorithms discussed don't appear to try to understand the behaviour of the markets in terms of the trading strategies of the market participants. Instead they try to predict future prices based purely on past price information. In my mind this is surely putting the cart before the horse, in the sense that in the markets, price and price changes are what result FROM the actions of market participants, not the other way round. Yes it is true that the actions of market participants are often driven by price, but it is always historical price information that is taken into consideration, even if it is a few milliseconds ago - the current market price is never precisely known until the order is placed and the trade confirmed.

 

The price at any moment in time is always the price at which buy and sell volume are exactly matched - if it is not, the price moves up and down in an instant (thanks to high-frequency traders) until it is exactly matched.

 

So if you could model the behaviour of market participants in terms of what volume they would each add (if they buy) or subtract (if they sell), then you could model the future of price changes.

 

Now of course this is very difficult because peoples' trading strategies are often complicated, chaotic, and subject to emotional influence, but it occurs to me that many new traders in particular are likely to be using simple trading strategies based on popular technical analysis methods, and similar. If we wanted to model this behaviour using AI we could potentially do this. And if this model was able to do this successfully then we would be in a better position to use AI to go on to model the prices that are more likely to occur as a result of this behaviour.

 

AI algorithms actually are just an application of technical analysis, but in a diiferent context.

'History can predict future events'; in tech speak, make the machine calculate all possible correlations in a data set - & it WILL find some that are 'somewhat' predictive (middling R-square values). As it applies these correlations, we call it 'learning'. Of course, the 'machine' is only as 'smart' as the R-square of the correlation, and it's stability in an out-of-sample application; introduce it to a market-discontinuity, and it goes beserk :)

 

One of the theoretical arguments around HFT is that if your holding period is very small (nano-seconds), almost all your price gain will be attributable to market drift; and we can calculate the amount of that drift, using the Brownian Motion equations. Applied to AI, the more you can apply the Brownian Motion equations to an AI algorithm, the more accurate and stable it becomes. 

 

All things coming out of the 'investment' silo, and making the jump into other places.

 

SD

...

???

I find the above comments quite interesting.

I've been using voice recognition software for quite some time and the technology relies on deep learning and machine learning, both subsets of artificial intelligence. Through recognition of voice patterns, the software reproduces written text and, over time, gets better at it. But the technology remains quite poor concerning certain aspects that require basic common sense (when I use a new word, a word in a different language, someone else speaks) and the "machine" does not recognize an obvious mistake with very potentially consequential impact on the substance of the underlying message. Proofreading has become markedly different as the software (even if amazingly efficient at certain tasks) can produce very stupid results.

 

neil9327's point, I think, was that we would hope to integrate and or understand the underlying "behavior" that led to the subsequent price action in order to improve prediction capabilities. The best short-term predictive ability of where a stock will go is what it did in the short-term past and this has been captured by simple linear regression models assuming markets function linearly most of the times (with some predictable variation) and this is where correlation coefficients and R-squared values come into play.

 

The idea (and the hope at this point) of machine learning and higher artificial neural networks for better prediction capabilities relies on improved pattern classification and ability to recognize patterns on its own in order to determine non-linear extrapolations.

 

In a way, this is nothing new as Thomas Bayes described the foundation of machine learning in 1763. Pattern recognition can be improved but the underlying principles that rest on past behavior can lead one astray (such as when using the VaR concepts) especially when transitions occur between calm and chaos or vice-versa. Artificial intelligence will need to integrate behavioral aspects and IMHO we're not quite there yet on many levels.

 

One of the biggest risks may be missing the forest for the trees (bigger picture, perspective etc) because the complexity of the model and the huge amount of data used may result in an illusory sense of precision.

 

I would say pattern recognition has value but is only a starting Brownian point for deep and independent thinking.

Potential bias: "Investment is most intelligent when it is most businesslike."

[Jurgis]

- I for one backflip for our robotic overlords

http://news.mit.edu/2019/mit-mini-cheetah-first-four-legged-robot-to-backflip-0304

[Jurgis]

https://www.sciencemag.org/news/2019/02/pictionary-playing-computer-connects-humans-deep-thoughts?r3f_986=https://www.google.com/ - Pictionarize this!

[SharperDingaan]

An interesting thread. One thing I notice is that the AI algorithms discussed don't appear to try to understand the behaviour of the markets in terms of the trading strategies of the market participants. Instead they try to predict future prices based purely on past price information. In my mind this is surely putting the cart before the horse, in the sense that in the markets, price and price changes are what result FROM the actions of market participants, not the other way round. Yes it is true that the actions of market participants are often driven by price, but it is always historical price information that is taken into consideration, even if it is a few milliseconds ago - the current market price is never precisely known until the order is placed and the trade confirmed.

 

The price at any moment in time is always the price at which buy and sell volume are exactly matched - if it is not, the price moves up and down in an instant (thanks to high-frequency traders) until it is exactly matched.

 

So if you could model the behaviour of market participants in terms of what volume they would each add (if they buy) or subtract (if they sell), then you could model the future of price changes.

 

Now of course this is very difficult because peoples' trading strategies are often complicated, chaotic, and subject to emotional influence, but it occurs to me that many new traders in particular are likely to be using simple trading strategies based on popular technical analysis methods, and similar. If we wanted to model this behaviour using AI we could potentially do this. And if this model was able to do this successfully then we would be in a better position to use AI to go on to model the prices that are more likely to occur as a result of this behaviour.

 

AI algorithms actually are just an application of technical analysis, but in a diiferent context.

'History can predict future events'; in tech speak, make the machine calculate all possible correlations in a data set - & it WILL find some that are 'somewhat' predictive (middling R-square values). As it applies these correlations, we call it 'learning'. Of course, the 'machine' is only as 'smart' as the R-square of the correlation, and it's stability in an out-of-sample application; introduce it to a market-discontinuity, and it goes beserk :)

 

One of the theoretical arguments around HFT is that if your holding period is very small (nano-seconds), almost all your price gain will be attributable to market drift; and we can calculate the amount of that drift, using the Brownian Motion equations. Applied to AI, the more you can apply the Brownian Motion equations to an AI algorithm, the more accurate and stable it becomes. 

 

All things coming out of the 'investment' silo, and making the jump into other places.

 

SD

...

???

I find the above comments quite interesting.

I've been using voice recognition software for quite some time and the technology relies on deep learning and machine learning, both subsets of artificial intelligence. Through recognition of voice patterns, the software reproduces written text and, over time, gets better at it. But the technology remains quite poor concerning certain aspects that require basic common sense (when I use a new word, a word in a different language, someone else speaks) and the "machine" does not recognize an obvious mistake with very potentially consequential impact on the substance of the underlying message. Proofreading has become markedly different as the software (even if amazingly efficient at certain tasks) can produce very stupid results.

 

neil9327's point, I think, was that we would hope to integrate and or understand the underlying "behavior" that led to the subsequent price action in order to improve prediction capabilities. The best short-term predictive ability of where a stock will go is what it did in the short-term past and this has been captured by simple linear regression models assuming markets function linearly most of the times (with some predictable variation) and this is where correlation coefficients and R-squared values come into play.

 

The idea (and the hope at this point) of machine learning and higher artificial neural networks for better prediction capabilities relies on improved pattern classification and ability to recognize patterns on its own in order to determine non-linear extrapolations.

 

In a way, this is nothing new as Thomas Bayes described the foundation of machine learning in 1763. Pattern recognition can be improved but the underlying principles that rest on past behavior can lead one astray (such as when using the VaR concepts) especially when transitions occur between calm and chaos or vice-versa. Artificial intelligence will need to integrate behavioral aspects and IMHO we're not quite there yet on many levels.

 

One of the biggest risks may be missing the forest for the trees (bigger picture, perspective etc) because the complexity of the model and the huge amount of data used may result in an illusory sense of precision.

 

I would say pattern recognition has value but is only a starting Brownian point for deep and independent thinking.

Potential bias: "Investment is most intelligent when it is most businesslike."

 

Good points.

 

The other issue with AI is that while development is happening in many different places, there's a lot of contagion as the many groups jump off each others innovations. While inherent to the scientific discovery, and agile project management, process; it produces 'dogma', along with discovery.

 

"We invented it, this is how you do it, don't presume to tell me otherwise".

At one time, we were also 'sure', that the earth was at the centre of the universe.

 

For AI to work commercially, it needs you and I to permit it 'free' access to large amounts of 'our' transaction data.

Whether at the granular, or meta-data level; that data is an asset - and you will be chaged to use it.

All learning has an ongoing 'tuition cost'.

 

To avoid spurious results, the data also has to be be complete - and accurate. Ever looked at historic data? It's full of inaccuracies,

Ever looked at blockchain data? Every transaction record is perfect and 2nd party verified - but you dont get it unless the Oracle makes it available to the public (distrubted/private ledgers). Oracles are the toll-booth trolls, and you WILL pay them to access their 'golden data'.

 

Blockchain/AI are opposit sides of the same coin ... but blockchain is the 'control' side of the coin.

Something that AI has been reluctant to recognize.

 

SD

 

[Jurgis]

http://fortune.com/longform/ai-drugs-pharma-pharnext-cmt/ - somewhat fluff piece, but covers the issues of going from gene-disease mapping to actual drugs that may treat a disease.

[Jurgis]

https://www.vox.com/future-perfect/2019/4/17/18301070/openai-greg-brockman-ilya-sutskever - Quo vadis Open AI.

[rkbabang]

‘You Can’t Take My Door’, A Country Song Created by a Neural Network That Studied a Catalog of Country Hits

https://laughingsquid.com/country-song-created-by-neural-network/

 

[rkbabang]

And even better.  I've found this.  A deep learning death metal generator.  It constantly streams deep learning generated death metal in real time as it generates it 24/7.  It isn't bad. Something you can just keep on your headphones as you work.

 

 

[Guest cherzeca]

in terms of AI used in equity investing, I am thinking that i) it is likely happening successfully now, as it seems Simon must be using some form of AI (or does Simon focus on other markets than equities?), and ii) AI pattern recognition is becoming profitably used in medical diagnostics (now tumor detection by AI is better than by a radiologist using eyes alone).

 

in a sense, the yes/no tumor detection decision is binary in the same sense as the gain/loss investment decision, but I wonder if the inputs are too multivariate in the investment context.  whether with go or chess, there are finite rules, although near infinite permutations. I wonder if there are infinite rules (let alone permutations) in equity investing...

[cameronfen]

in terms of AI used in equity investing, I am thinking that i) it is likely happening successfully now, as it seems Simon must be using some form of AI (or does Simon focus on other markets than equities?), and ii) AI pattern recognition is becoming profitably used in medical diagnostics (now tumor detection by AI is better than by a radiologist using eyes alone).

 

in a sense, the yes/no tumor detection decision is binary in the same sense as the gain/loss investment decision, but I wonder if the inputs are too multivariate in the investment context.  whether with go or chess, there are finite rules, although near infinite permutations. I wonder if there are infinite rules (let alone permutations) in equity investing...

 

Simon focuses on currency markets I think.  Most of his stuff is trend following I think and is definitely uses at least some AI since I know he hires top ML Phds from MIT and the like. 

[SHDL]

in terms of AI used in equity investing, I am thinking that i) it is likely happening successfully now, as it seems Simon must be using some form of AI (or does Simon focus on other markets than equities?), and ii) AI pattern recognition is becoming profitably used in medical diagnostics (now tumor detection by AI is better than by a radiologist using eyes alone).

 

in a sense, the yes/no tumor detection decision is binary in the same sense as the gain/loss investment decision, but I wonder if the inputs are too multivariate in the investment context.  whether with go or chess, there are finite rules, although near infinite permutations. I wonder if there are infinite rules (let alone permutations) in equity investing...

 

The story is that this is what Renaissance Technologies has been doing since inception.  From what I know they seem to focus mainly on relatively small short term trades where they can exploit whatever patterns or anomalies they find in the data.  They then milk each trading strategy until it stops working, at which point they move on to something else. 

 

I think it’s much harder to do this sort of thing as, say, a long term buy-and-hold equity investor because of data limitations and the fact that statistical relationships between financial/economic variables may well change over the course of several decades. 

[Guest cherzeca]

"statistical relationships between financial/economic variables may well change over the course of several decades."

 

several decades, or several weeks...?

 

there is another issue as I think about it, and that is that humans are cause--->effect creation machines, we need a narrative to explain what otherwise is chaos, and so we say that this causes that when, in the social/economic/political world outside a chemistry lab, we may be satisfying a psychological need more than defining any objective relationship.  (mauboussin writes about this).  perhaps the best advantage of AI is for it to be more conjectural about cause and effect, and more able to  continually refine

[SHDL]

several decades, or several weeks...?

 

I did mean several decades.

 

For example if an anomaly exists just because there is some idiot involved in a certain name, it may be reasonable to expect that to continue for a few weeks but probably not for decades. 

[Guest cherzeca]

several decades, or several weeks...?

 

I did mean several decades.

 

For example if an anomaly exists just because there is some idiot involved in a certain name, it may be reasonable to expect that to continue for a few weeks but probably not for decades.

 

I guess my point was that if AI is going to detect patterns it would be useful if those patterns persist; and in equity markets I am not sure the there are long term persistent patterns

[SHDL]

several decades, or several weeks...?

 

I did mean several decades.

 

For example if an anomaly exists just because there is some idiot involved in a certain name, it may be reasonable to expect that to continue for a few weeks but probably not for decades.

 

I guess my point was that if AI is going to detect patterns it would be useful if those patterns persist; and in equity markets I am not sure the there are long term persistent patterns

 

I guess we share the same skepticism then.

 

I do think some patterns will survive forever though - like if you buy cheap you should do better all else equal. That’s one of the big reasons I believe in value investing.

[Liberty]

Really interesting interview with Geoffrey Hinton on neural networks:

 

https://www.wired.com/story/ai-pioneer-explains-evolution-neural-networks/

[UNF2007]

We got an email from the CEO of the hospital system I work for, that they are implementing an AI program, to scan all charts and report if anyone is accessing medical records that they shouldn't. Right now they use a random audit system, but supposedly the AI will be monitoring every person on the system in real time.

[cameronfen]

We got an email from the CEO of the hospital system I work for, that they are implementing an AI program, to scan all charts and report if anyone is accessing medical records that they shouldn't. Right now they use a random audit system, but supposedly the AI will be monitoring every person on the system in real time.

 

This sounds like it doesn't need AI, but maybe I don't know.  Shouldn't you just write a series of if-then statements to check if someone's account is accessing records they are they don't need?  Sounds like they are using fancy words to scare people.   

[Jurgis]

We got an email from the CEO of the hospital system I work for, that they are implementing an AI program, to scan all charts and report if anyone is accessing medical records that they shouldn't. Right now they use a random audit system, but supposedly the AI will be monitoring every person on the system in real time.

 

This sounds like it doesn't need AI, but maybe I don't know.  Shouldn't you just write a series of if-then statements to check if someone's account is accessing records they are they don't need?  Sounds like they are using fancy words to scare people. 

 

It's not as trivial as you suggest.

 

For clear cases, you just have access control, you don't need if-then statements.

But there's a lot of situations where access control does not work and if-then statements don't either.

 

Is the nurse accessing patients' records five times a day checking that his temperature is OK or is she checking out a sexy guy's disease info and telephone number?

Is a specialist accessing data of a patient that they saw three days ago checking whether they correctly entered the info or are they trying to preempt some error/negligence they made?

 

You can't just access control these, and you can't just flag them all since you gonna get too many false positives.

 

So you run fuzzy, statistical systems on the data. And you can call them AI since it is really what AI has been for ages. It might not be DNN, might not even be ML, but sure it's AI. BTW, when you say "if-then statements" - if it's a rule based system, that's AI too.  ;)

And this stuff has been used for similar use cases - like fraud detection - since 1990s if not before.

 

(In general, if it's in Russell and Norvig, then it's AI https://smile.amazon.com/dp/B00I2XV9IY/ref=cm_sw_em_r_mt_dp_U_ntK3CbQ0M8RP3  8) )

[Castanza]

BTW, when you say "if-then statements" - if it's a rule based system, that's AI too.  ;)

 

That's not AI. That's a static deterministic model using ML algorithms within a bound set of inputs. AI can take inputs outside of the bounds and return a probability factor based on confidence. If-then or if-else based models are purely deterministic and only operate within the bounds of inputs determined by the programmers. Boiled down it's basically a decision tree with integrated ML. If you want to call that AI, fine haha but it's not "true" AI.

[Jurgis]

https://en.wikipedia.org/wiki/Expert_system

[Castanza]

https://en.wikipedia.org/wiki/Expert_system

 

"However, some experts point out that expert systems were not part of true artificial intelligence since they lack the ability to learn autonomously from external data."

 

 

[rkbabang]

https://en.wikipedia.org/wiki/Expert_system

 

"However, some experts point out that expert systems were not part of true artificial intelligence since they lack the ability to learn autonomously from external data."

 

 

It's funny that whenever we get a system to do things that computers could never do before we declare that it isn't "real" AI and we move the goal post further back.  To some "real" AI won't be achieved until artificial systems can think, think about thinking, feel, love, hate, fear, not want to die ....

[Castanza]

https://en.wikipedia.org/wiki/Expert_system

 

"However, some experts point out that expert systems were not part of true artificial intelligence since they lack the ability to learn autonomously from external data."

 

 

It's funny that whenever we get a system to do things that computers could never do before we declare that it isn't "real" AI and we move the goal post further back.  To some "real" AI won't be achieved until artificial systems can think, think about thinking, feel, love, hate, fear, not want to die ....

 

I think it's more people are skewing what AI really is making factitious claims. It's no different than these companies that claim they've built an atomic computer. I don't think the definition of AI has ever changed and I don't think we should change it to meet technological short comings. All we are doing is processing more data faster. It can give the appearance of Intelligence, but is it truly intelligent?