Should Validation Differ with Machine Learning?
HTML-код
- Опубликовано: 30 мар 2024
- Should Validation Differ with Machine Learning? No!
Machine learning is nothing more than a small space of statistics. We are building models on the same data sets to model the same problems. Why would the process change for this? Why would you blindly use a method and do minimal or no testing? I ask these because the machine learning and data science community has been trying to convince people that ML models are a magic pill to fix everything. It doesn't change much besides a new approach with the same data and business problem restrictions.
For validations make sure to review the data sampling and cleaning, conceptual soundness, model performance, and model governance.
Model Validation: Detailed Process
• Model Validation: Deta...
Agus Sudjianto Interview
• Machine Learning in Fi...
Decision Tree Number Space Details
• Decision Tree Number S...
Website:
www.FancyQuantNation.com
Support:
ko-fi.com/fancyquant
Quant t-shirts, mugs, and hoodies:
www.teespring.com/stores/fanc...
Connect with me:
/ dimitri-bianco
/ dimitribianco
Sure I’ll watch this at 4:00 am
Lol, you must be on the west coast.
@@DimitriBianco sure am
It would be great to see a video discussing measure theory and its applications in specific ML models. I still haven't found a resource directly linking measure spaces and ML models.
Start here:
ruclips.net/video/S1Yyn9Vs5Tk/видео.htmlsi=13vKjJiC0KLb1V1g
I'll consider this for future videos.
I've been having success using GBDTs for OOS/OOT predictions for adjustments to existing financial models to better reflect the idiosyncrasies of a more recent market/rate environment. However, as you say, you need to be especially careful to ensure the model is sufficiently simple and stable for it to be usable. And, again, this model is only predicting relatively small but still material adjustments to an existing set of pricing models. I've also been careful to prescribe the limitations of the model, as you mention (e.g. we dont have a sufficient number of samples with high x and low y to provide a meaningful prediction or we can expect more flat residuals across feature A, but inversely scaling residuals across feature B). The model has also been good at helping to visualize changes across certain features in time as the moving-window of prediction has changed (e.g. specific counterparty behavior in certain OTC markets relative to the market as a whole)
That's really interesting.
Would love to know your oppinion on neural SDE, SPDE learning or Physics Informed Neural Networks (PINNs) that can either learn a solution space given a mathematical restrictions or a model like a SDE, or learn the SDE and their solutions directly from the data.
Good stuff bro!
I am a quant and 2/3 failure rate is honestly pretty good 🤣 I think 1/10 or fewer of what I come across works to our standards
I was hated by most developers especially when I would make them go down the math path and explain all the issues.
Can you explain what you mean by number spaces ?
i think in mathematical terms, he means domain or pre-image/inverse image
Here is a detailed video on it. I also have a simple video which is more straight to the point if you search.
ruclips.net/video/S1Yyn9Vs5Tk/видео.htmlsi=vJLe_UET5GY9ZSgt
hi dimitri, can any one who has started their career as quant researcher in HFT like Tower Research etc. can make transition into buy side firms like goldman, jp morgan etc after working in hft for 5-6 yrs?
Please reply
Thanks
Investing like HFT is considered buy side and the banks are considered sell side. Yes, it is possible to transition but as you specialize it can become more difficult.
@@DimitriBianco Thanks For Your Reply
Nice gaming pc. Or is that a mini local AI/ML rig?
That's actually my work pc.
Tikhonov regularization 😉