Machine learning in trading: theory, models, practice and algo-trading - page 3743
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
The more I delve into LeCun’s ideas, the more I agree with him that no AGI will be created on the basis of LLMs. His approach offers some cause for hope, but only time will tell.
“Intelligence is not what you know, it’s what you do when you don’t know.” – from his lecture:Very often, my actions precede my thoughts
then you sit there and try to explain to yourself why you did it that way, and you have to think really hard ;))))
I’ve finally understood, thanks to the quote being translated in the post above
But I thought it wasn’t intelligence, but intuition
In other words, the action takes place at the level of intuitionObviously.
It’s not intelligence, but intuition
In my humble opinion, one doesn’t get in the way of the other at all.
It’s one thing to know something and quite another to be aware of that knowledge. One could argue that intuition is knowledge without being aware that one possesses it.
Intelligence is not what you know, it is what you do when you do not know.
‘Intelligence is not what you know, but what you do when you don’t know’ — this saying is often attributed to the Swiss psychologist Jean Piaget, who studied the development of children’s thinking. Even if the original quote sounded slightly different, the phrase itself is wonderful in that it points to a definition of intelligence that is inextricably linked to the concept of purpose. I asked the AI to draft a short essay on this topic:
The Teleology of the Mind: Why Intelligence Is Measured Solely by Purpose
In its attempts to define intelligence, science often reaches an impasse, describing it as ‘the ability to process information’. However, this definition is rather meaningless. Any calculator, thermostat or sheet of paper on which ink is drying technically processes information, converting input signals into output. Such an approach blurs the boundaries of the term and, crucially, deprives us of the ability to objectively measure a system’s rationality.
The most profound, rigorous and operational definition is the teleologically oriented one: intelligence is a system’s ability to construct, restructure and effectively implement a hierarchy of goals under conditions of uncertainty.
The scale of measurability: from a calculator to a human being
It is precisely this link to the concept of a goal that transforms abstract philosophical speculation into a precise metric. Intelligence becomes measurable because the degree of its development is directly proportional to the complexity of the tasks the system is capable of solving. This allows us to construct a single, continuous scale of the evolution of the mind:
Conclusion
Intelligence is not a passive process of ‘chewing over’ data, but an active expansion of the will into the environment. By defining intelligence in terms of goal-directedness, we obtain a universal assessment tool. We can measure intelligence mathematically — as the efficiency of maximising an objective function across a wide range of complex environments. Without a goal vector, intelligence is blind, immeasurable and indistinguishable from the background noise of the universe.
I asked the AI to draft a short piece on how mathematics dashes our hopes of an imminent AGI. This certainly won’t help against those who predict it, but here goes:
A mathematical impasse: the continuum of environments versus algorithms
If we digitise the infinite stream of environmental inputs and the system’s responses as integers, then any ‘correct’ behaviour is a function from integers to integers. A simple fact is known from mathematics: the number of all possible functions has the cardinality of the continuum, whilst the number of any conceivable algorithms (programmes, neural network weights) is strictly countable. Almost all theoretically possible adaptation functions are, in principle, non-algorithmisable.
It follows from this that the ‘hallucinations’ of modern AI chatbots are inevitable. A neural network is a finite, static algorithm. Should it venture beyond the domain for which it was trained, it is mathematically bound to return an error, as it is physically incapable of containing trajectories for the infinite variety of the environment.
Games of reflection: why heuristics are not the answer
One might assume that to create strong AI (AGI), it is sufficient to find working ‘heuristics’ – rough simplifications of the environment. However, situations modelled by game theory (such as stock market trading, military strategy or geopolitics) prove this assumption to be false. In environments characterised by active opposition, any ‘correct’ algorithms instantly negate their own correctness due to reflection. As soon as a system figures out the market, other players figure out the system and change the rules of the game.
This means that it is fundamentally impossible to create AGI in the form of any fixed algorithm, however complex it may be. True intelligence is not a set of ready-made heuristics, but a continuous meta-process. It must remain in a state of constant replanning, where the previous ‘correct’ model is erased as soon as the environment completes a reflexive cycle.
This is an interesting paper : https://arxiv.org/pdf/2304.07619
Our key finding is that GPT-4 demonstrates a strong ability to predict the immediate market reaction to news headlines, achieving daily portfolio hit rates of approximately 90% for correctly identifying the direction of the initial price response. This result is notable for two reasons. First, it is not obvious that a general-purpose language model trained without explicit financial supervision should excel at this task. The fact that it does suggests that sophisticated financial reasoning—understanding not just whether news is positive or negative, but also how markets will interpret it—emerges naturally from the model’s general language training.1 Second, and more importantly for financial economics, this high accuracy provides a unique window into market information processing.
achieving a daily forecast accuracy of around 90% for the portfolio, provided the direction of the initial price reaction is correctly identified.
I’ve had a read. It’s cool and interesting, but the content of the news might not actually be relevant. The LLM could have picked up on the volatility forecast simply based on the number and variety of news items.
There’s a lot of news and it’s all over the place, so volatility rises; when there’s little news or it’s mostly the same or consistent, volatility is low.
High volatility signals a reversal, whilst low volatility signals a continuation. And this tends to last for about 1–2 days.
Here’s an interesting article: https://arxiv.org/pdf/2304.07619
They’ve realised a bit too late. From there:
Our theoretical framework posits that LLMs can increase investors’ information processing capacity and reduce market inefficiencies. As a result, return predictability should weaken as LLMs grow in size and more investors begin to use them. Whilst testing this hypothesis presents an empirical challenge, we find some suggestive evidence: a general decline in the performance of the ChatGPT-based strategy during our sample period, over which the capabilities and adoption of GPT models skyrocketed. For example, its annualised Sharpe ratio falls from 6.54 in Q4 2021 to 3.68 in 2022, 2.33 in 2023, and to 1.22 over January–May 2024.
its annual Sharpe ratio fell from 6.54
6.54 – really?!
Well, that’s not a question for me – the authors’ contact details are given in the footnote on the first page.
I’m not sure how feasible a backtest would be to verify this – you’d need the same data, model and trading system.