2. • Big promises of quantitative
finance and even bigger promises
of ML in finance
• Reality in academia, industry and
practice
• Making ML work with financial data:
hands-on improvements
• Roadmap: how to build and keep
that trust in the relationship
Agenda
https://towardsdatascience.com/ai-in-
finance-how-to-finally-start-to-believe-
your-backtests-1-3-1613ad81ea44
3. Alex Honchar
• Co-founder and Chief AI Officer at AI
solutions firm Neurons Lab
• AI practitioner for the last 7 years,
SMBs and startups, fintech and
medtech
• Educator, University of Verona
professor, 1M+ views at Medium
blog, 170+ research papers
quotations
About me
4. Big promises of quantitative finance
Hundreds of years and the same formula
Carl Friedrich Gauss, 1809 William Sharpe, 1964 Andrew Ng, 2010s
5. Big promises of quantitative finance
Linear regression as the panacea
• “If something grows or falls, the
asset price will react with growth or
fall as well”
• Correlation, linear regression, factor
modeling - they all measure the same
thing
• Modern “somethings” have multiple
variables, but relationship stays the
same: linear
CAPM: https://www.educba.com/capm-formula/
https://corporatefinanceinstitute.com/
resources/knowledge/finance/fama-french-
three-factor-model/
6. Big promises of quantitative finance
What to do with total randomness?
• “If both somethings and the asset
price are following random
distributions, let’s put it in the formula
too!
• Geometric Brownian Motion, Jump-
Diffusion model, Stochastic
Volatility…
• We see the pattern in the past,
explain it (by hands) in the past,
evaluate it in the past
https://
quant.stackexchange.c
om/questions/32763/
will-volatility-
smoothing-effects-
exist-for-returns-driven-
by-geometric-brownian
A small addition to GBM:
https://en.wikipedia.org/wiki/Stochastic_volatility
7. Big promises of quantitative finance
Alright, but what’s wrong?
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3373116
9. Even bigger promises of AI
Let it all “learn by itself”
• Potentially AI&ML promises to
revolutionize everything based on the
successes in other fields
• Factors? Boosting models win in
Kaggle!
• Stochastics? Recurrent nets do it
the best!
• Patterns? Convolutional nets do it
themselves!
11. Real-world exercise
You can check out the code in my blog
• AAPL market and fundamental data
• Well-split and normalized
• MLP neural network to predict the price in
the future
12. Real-world exercise
What could be done wrong?
• Perfect time series fit!
R^2 ~ 1.0, MAE ~ $4!
• ML strategy horribly
fails compared to the
HODL benchmark :(
24. Okay, but this is enough, right?
We wish it was :)
• Where did you get the data from? Yahoo Finance? Like millions of others?
• Where is cross-validation for the ML part?
• Single scenario backtesting? Overfitting to the past again?
• How did you guess those parameters? Randomly, eh?
• All those tweaks, how many times you did it? Isn’t it multiple comparison
pitfall?
25. Next steps
How to step up your game ASAP
• Focus on the data:
• Remove the noise from the inputs and the outputs
• Make them practically and financially appealing
• Models don’t matter, what matters is that how much you can trust them:
• Feature importance is a king
• Do the right cross-validation
• Backtesting is the very last thing you want to do! Even if you arrive there, check for
multiple comparison and multiple scenarios, not a single historical backtest!