Thread by @HangukQuant

(14) 1. Throwback: Remember interviewing for quant firm and when explaining projects I talked about AI models I built. We discussed some diagnostics about where the models failed. PM took a look and said, "well garbage in garbage out"😳 thread on application of AI in markets. 👇

2. This is a #travelthought, so bear with me if some points are incoherent. First, my advice on approaching AI

https://twitter.com/HangukQuant/status/1441889665850626048?s=19

Now, AI sounds cool, so this warning obviously is not enough to put you off. If you insist on going down that path, then let's dive in.

3. First, let's take it down from its pedestal and demistify this buzzword. It is nothing more than a class of search algorithms.
Brute force? Let's improve. B/DFS? We can do better. A*? Now what if our task setting is partially observable?

4. Artiticial Intelligence problems are not created equal. Uncertainty, stochasticity, adversarial, multi-agency, dynamism and continuity are all environmental factors of task settings that make AI problems difficult to "solve". What is "solving"?

5. Solving a problem is finding an optimal/true solution among candidate solutions in a hypothesis space. For example, y = b + mx. Parameters (b,m) form candidate solutions in 2D space. But these are generalised into higher dimensional Euclidean spaces when parameters increase.

6. So we establish that AI is a search problem. The "Intelligence" comes from the heuristics and guidance behind the algorithms approximating some optima that traverses search spaces in efficient ways, using ideas behind differential calculus and iterative numerics.

7. First, you have to understand the mathematics behind these algorithms. That involves understanding conditions for convergence, and assumptions for accuracy of output. How can the output of a model be fed into another as input? Are there issues when assumptions are stacked?

8. I have yet to even comment much on nature of markets. AI itself is a vastly complex community of thinkers, and tapping into them require understanding some subtleties. Most importantly, you need to understand assumptions that accuracy relies on.

9. For example, take linear regression. What are the assumptions for model fitting? Constant variance, independent predictors, independent errors. What happens when there are error correlation? What happens under multicollinearity. Understand individual techniques.

10. Next step is to understand comparison. Why do linear regression and when you can do KNN regression? Their errors scale differently with noise. Etc. What about neural nets as functional approximators? You get the point.

11. Now we have finance/trading, which brings itself a host of problems. Lets call them "stylized facts", observations that derived from empiricity. One of such facts is that in trading/alpha the noise to signal ratio is much larger than in most other domains.

12. To make matters worse, our optima/goal is dynamic and alpha is restless. So what algorithms tend towards such settings? Which are more robust to noise? Which problems in trading are easier to approximate? For instance, "risk" is easier to approximate than "returns".

13. That means there are different problems within trading, different markets within trading, different techniques in AI and a careful selection of them need to emsembled to tackle different problems. Engineering coherent solutions require analysis of both domain and technicality

14. AI is not a magic solution, but it is also not a good-for-nothing buzzword. A bad workman always blames his tools.
Fin. Happy trading!