Unlocking the Mystery: Why Stock Markets Are Unpredictable 📈

Hello Everyone,

In this episode of The Casual Causal Talk, I had the privilege of speaking with Dr. Stephan Kolassa ( / stephankolassa ) - a leading expert in Time Series Forecasting with deep experience in retail forecasting.

Back in the pre-ChatGPT/Perplexity days, if you had questions about statistics, forecasting, or ML, Cross Validated was the go-to place.

Among all the answers, one voice always stood out for its clarity, rigor, and depth, that of Dr. Kolassa himself.

For years, I have wanted to pick his brain on forecasting, and this conversation finally gave me that chance.

And what a conversation it turned out to be !!

Time Series Forecasting isn’t just academic, it is everywhere: Marketing Mix Modeling (MMM), Retail Forecasting, Inventory, Supply Chain, Finance and even in how LLMs work (yes, there is a autoregressive element. More on this in podcast). Also temporality is one of the core tenets to establish causality.

This episode goes beyond a podcast, it is essentially a masterclass. I was fortunate to get a front row seat, and now you can too.

We covered a wide range of fascinating topics and Dr. Kolassa has also curated a set of resources (pls see the first comment).

Tune in, learn, and enjoy this insightful deep dive into the world of forecasting.

00:00 - Introduction

00:30 - Don't trust LLMs, Trust Cross Validated

02:25 - Stephan's Career Journey. How a psychologist (Stephan's wife) turned a mathematician into a statistician.

04:24 - Forecasting with importance to inference.

05:25 - We are living on a time arrow

06:00 - How time matters in domains like credit scoring and image classification

07:50 - What is time series forecasting ?

09:00 - Retail forecasting and its importance

10:00 - Components of time series

12:20 - What is smoothing exactly? Why is it even called so? What is exponential smoothing?

15:00 - The Makridakis 'M' competitions

16:40 - The humble exponential smoothing outperformed ARIMA in M1 competition !!

17:25 - 92.5 % submission didn't outperform exponential smoothing in M5 competition where retail data - Walmart retail data was used.

19:00 - Exponential smoothing beats complex forecasting methods often

21:20 - The reason why Exponential smoothing beats complex forecasting methods - Bias / Variance trade off?

24:00 - Is ARIMA the hello world of time series forecasting?

26:45 - Understanding the AR, I and MA components of ARIMA.

30:00 - How to find the optimal parameters for ARIMA model

31:15 - ARIMA is popular because it helps prove things and mathematicians / statisticians love proving things.

33:42 - How to validate ARIMA with Auto ARIMA.

34:50 - Should the number of iterations for Auto ARIMA be increased to give it a fair chance to recover the parameters of ARIMA model.

35:50 - Autoregression is present from time series to LLMs - why it is so effective and does it have some form of memory encoding?

37:06 - When Autoregression can fail and how it can affect causality.

39:00 - In retail forecasting, mere forecasting is not enough, explain-ability is important too.

41:00 - Marketing Mix Modeling (MMM) blends best of both worlds - Forecasting and Explanation.

42.00 - What is Seasonality? Does ARIMA handle it well or SARIMA a better option?

48:00 - Is more data better for forecasting? Examples from MMM and Retail Forecasting.

51:00 - How to navigate periods like Covid for time series forecasting?

51:40 - Forecasting is not just science it is art as well.

52:00 - Why dummy variable encoding is bad and how RBF methods are better

55:35 - How RBF helped Aryma Labs increase accuracy of its MMM models

56:45 - What is hierarchical forecasting?

01:02:00 - Hierarchical reconciliations.

01:04:00 - Quantile Forecasts

01:06:00 - Is it a good idea to model at a granular time level like seconds, minutes or hours?

01:09:00 - Why K fold validation is bad for validating time series forecasting ?

01:11:00 - The difference between prediction intervals and confidence intervals

01:13:05 - What are conformal predictions?

01:18:00 - Are MAPE, MASE and RMSE inadequate?

01:26:00 - What is forecast-ability ? What can be forecasted and what can't be?

01:31:00 - Don't just forecast. Forecast for business decisions

01:33:00 - Why stock markets can't be predicted - the efficient market theory.

01:34:00 - Could there be causal forecasting?

01:36:25 - Are ensemble methods in forecasting underrated or overrated ?

01:37:30 - What is the first thing a company should do to get forecasting right?

01:39:13 - Success stories of forecasting from industry

01:42:40 - Resources to get started on time series forecasting

01:44:00 - Conclusion