Hi @PsychedelicStick I worked in investment bank, hedge fund and sovereign wealth fund. What makes the job different is the time scale :
IB are intermediaries: very short term (Value at Risk computed at 1day), they commit capital and engage all the risk they can (under liquidity and regulatory constraints)
HF are taking medium term risk (1 day to 2 weeks): they commit risk because what they maintain is a Sharpe ratio (ie more risk => more returns), and they engage all the capital they can (via prime brokers that can offer leverage if the HF strategies are diversifying enough)
the SWF are very long term and have no principal-agent problem, it means they start by securing some capital/liquidity that they need as principal, and engage all the rest in diversification (for what it means regarding the economy of their country).
Same modelling capabilities, same datasets, but three different time scales and hence very different focus.
In my personal case in the IB I was focusing on liquidity and execution, in HF in launching new programs (including alternative data for systematic investment), and in SWF in rethinking the global strategy around the rise of ML and data sciences.
About your second point (the transformation of data analyst jobs).
I do not think they will disappear, in some sense it will be the opposite:
data analysts will have more tools for exploration and in-depth statistical analysts
the biases in datasets will continue to be very important, and only human economic thinking will allow to identify them (as of today the optimal correction of the bias is an open question)
At the extreme, you could even state the opposite of your statement: if vibe coding and chain of thoughts for pipeline deployment becomes cheap, the only added value will be choosing the datasets! This is where the decision is taken. In such a world one will only need
a portfolio analyst to understand the weaknesses of the portfolio
discussing with a data analyst to choose a dataset than may contain the information that is needed to correct these weaknesses!
Difficult to say. In research PyTorch is more used, but in production TensorFlow is present.
How to understand that?
is it just a slow migration from TF –> PT happening, and will Jax be the next one?
is it the never-ending “explore fast and easy –> deploy on robust architecture” duality?
is it that GCP is more suitable to deploy and monitor TF than any other cloud can support the deployment in prod of PT?
It is the occasion to write I was a big fan of TFX (TensorFlow eXtended) and I have not seen something similar for PyTorch. I developed an interest for AAD in Finance long ago (MINI-SYMPOSIUM ON AUTOMATIC DIFFERENTIATION AND ITS APPLICATIONS IN THE FINANCIAL INDUSTRY), it came from the first implementations of Torch, long long ago now. So I am in between: no personal preference between the two.
Very good point @KKR444 , it should. The question it to teach what you know, or what you know enough to gather a good group of teachers around you.
Moreover, there is a deep connection between maths finance (that is the base of quant finance): stochastic control.
That is what you use to build portfolio and hedge derivatives.
That is also what drives the stochastic gradient decent that is the base of how artificial neural networks (ANN) learn: you control the trajectory of the weights on a random sample of points until the ANN performs the tasks described by the Los function.
This there is a ground to teach simultaneously ML and quant finance: there is a way to understand better one while you learn the other (this goes both ways). That is what we want to teach to out students.
A while ago I has long directions with one. This is an option that it a mix of startup up and “buy instead of build” logic. Interesting, depending on where you are in your career; you probably should have been is a very hands on team before (in modeling, estimation and coding).
In general, I have a tendency to think it is better to have been in the shoes of an agent before being the principal.