Time Series Forecasting with Machine Learning

INVESTING
[1] Webull (You can get 3 free stocks setting up a webull account today): a.webull.com/8XVa1znjYxio6ESdff
TIMESTAMPS
0:00 Introduction
1:51 Defining Problem
2:50 Understanding the Data
3:18 Analyzing Data (Trend, Seasonality)
4:40 Traditional Timeseries Forecasting (ARIMA, Prophet)
6:01 Univariate & Multivariate Time series
8:15 Time series with Machine Learning
9:02 Types of Time series models
11:05 Machine Learning Vs. Traditional Time Series
REFERENCES
[1] Math behind Facebook prophet: / the-math-of-prophet
[2] Traditional time series analysis step by step: www.kaggle.com/freespirit08/t...
[2] Dealing with time series data: online.stat.psu.edu/stat510/l...
[3] Catboost is slick : catboost.ai/docs/concepts/tut...

[@chinmayeejoshi4592]

I just want to say how much I love that it’s my grandma that has a laptop repair shop. 😍

[@krzysztof1156]

This is such a great video, I'm in the middle of writing my master thesis where I use timeseries and it helps a lot to understand the concepts.

[@vincentguttmann2231]

I was looking for an introduction to time series forecasting for a personal project (gas price prediction, since gas prices here in Germany are kinda high), and this was the perfect primer for time-series forecasting. Not too dumbed down, and not too complicated. And, obviously a GREAT example.

[@CodeEmporium]

Thank you! So glad this helped!

[@jasonsykes4199]

Don't stop making videos. You have a great teaching video.

[@myal7532]

Thank you for this awesome video! I'm pretty new to ML and time series and this is so helpful and clear. I'm actually working on an assessment for a Data Analyst role that I'm interviewing for and I'm tasked with forecasting travel bookings. Glad I came across your video and excited to check out your other ones!

[@zatarawood3588]

Did you get into a data analyst role?

[@sauravdas9751]

This video is great! I loved how you have put down the different methods so clearly and their pros and cons. Cheers to more videos!! :)

[@joshuabradshaw1647]

Thanks for the awesome comparison! Very insightful!

[@hoangphuoc3223]

wow. you explain concepts very well

[@joguns8257]

Very simplified. Thank you.

[@abhijitroy99]

Simple explanation and very good

[@sayantanghosh6714]

Thank you for the useful explanations!

[@ErturkKadir]

Great explanation. The big challenge of some time series is changing characteristic over time, which makes the prediction impossible. Like some changes appears and disappears randomly, no seasonal features. Plus sudden changes. What grandma could do if all of sudden thing changes, and this happens frequently. All of them makes the problem interesting.

[@chrisjfox8715]

Uber's algorithm uses an LSTM Autoencoder NN tuned via tons of time series training data, then takes the latent vector from the time series test data and feeds it [concatenated with more current conditions like today's weather and impromptu events] into a Perceptron NN. Those impromptu/external features are what's key.

[@RezaZen]

Great explanation!

[@shakirullah5840]

Thank you so much for your nice video. -- From Bangladesh

[@sanjaisrao484]

Thank you for amazing explaination

[@parakhchaudhary7479]

This is a seriously great introduction!

[@CodeEmporium]

Thanks a ton!

[@abhishekk1231]

Loved this video...! Thankyouuuuuu

[@plenilune4975]

Super helpful! Thank you so much!

[@IdiotDeveloper]

Really nice video.

[@huzifighter9]

Hi, great video! I'm just getting into time series forecasting, and you teached me a lot, thank you :) Could you make a video about Graph Attention Networks for time series forecasting?

[@frytka12345]

One note: you can actually use the machine learning model (the non-traditional model as you call it in the video) to dynamically predict whatever number of future points you want, you just have to implement recursion manually. Train the model to predict one step ahead, then use that prediction to predict 2nd step and so forth. This will very likely become "hard to get right" similarly to what you said about traditional models, as it's a much more complicated problem, but it is doable.

[@amirghorbani7922]

It is better to use icdst Ai predict lstm model.

[@frytka12345]

@@amirghorbani7922 There's no model that is best among all of them. Try many and choose the best one

[@vk7184]

Thanks a lot, very useful for me. I was wondering whether we can use time series forecasting using regression trees or not?

[@Shoaibkhan-oj3oe]

I loved this video. Such great information easily explained. Thankyou

[@CodeEmporium]

Thank you!

[@Arjun_Adapalli]

Thank you for this video!

[@Sanjeevsachin24]

very well explined.. the issue was also looked at the business front as well.. were as the traditional IT gig would explain in the point unlike this.. it shows the understing of the business is important to adapt to these new tecnologies..

[@CodeEmporium]

Thank you! Glad this perspective us helpful

[@fullduckdev7327]

thanks. very informative!

[@moreisdifferent]

Thanks a lot for this video! It gave me a perspective I badly needed. Two comments about it, one about the content and one about presentation: I wish you talked a little bit about how uncertainty quantification works in traditional vs ML approaches. For example, when you compare extendability of traditional vs ML models, I can imagine each round of recursion in the traditional models must account to some added error that we need to keep track of, and it would be nice to know how hard a problem that would be. I can imagine you kept this introduction video lean to be digestible, but I get anxious when forecasting is discussed without a mention of uncertainty... And the one about presentation (and it may be only me with this problem): the animation you use on text popping up, distracts me. I want to read it right away, and I have to wait a bit, and it 's mildly irritating. Again, it may be just me. Thanks again, and looking forward to your next videos on time series.

[@muritalaadebayoisah9155]

So precise!!! THanks

[@chuckcheddar461]

What if you want to output a vector? i.e. you want the model to output laptops you will receive on each day from 1-10 days into the future (output vector of size 1x10). Great video!

[@user-rq8uv6ir9w]

You are awesome!! Thank you for this video

[@CodeEmporium]

You are extra welcome :) Thank you for supporting

[@henriquepalhoto2257]

good stuff

[@nourinahmedeka9518]

This was awesome!

[@sgrouge]

I worked hard on forex dataseries: EURUSD tick resolution, compressed with wavelets, passed into LSTM under keras. Got 73% accuracy on the next minute: not bad for experimental results? What gives me headaches: - do I always need to make the timeseries *stationary*? - How to scale perfectly my timeseries, according to what model im going to use (lstm, mlp, sklearn regressor...)? - Do I have to use stateless or stateful lstm??? - Does it have soem sense to shuffle sequenses before training lstm? I could not find clear answer anywhere on the net...

[@valdrich472]

Amazing timing! I am just getting started with time series forecasting. Do you have any tools/sources to follow? Thanks man! Love your videos!

[@AndyTutify]

Check out Rob Hyndman's Forecasting: Principles and Practice. It is freely available online.

[@DistortedV12]

This is still rather opaque..for certain classes of machine learning models it seems like the data assumption is i.i.d correct? Such that datapoints farther in the time horizon will be treated exactly the same as datapoints closer to the time horizon; maybe I am not understanding.

[@prometeo34]

Great explanation, thanks

[@CodeEmporium]

Super welcome!

[@RayTayek]

nice overview. thanks.

[@CodeEmporium]

My pleasure. Hope you like the rest of the channel :)

[@tony7682]

Please correct if wrong: Traditional time series models are not necessarily recursive: MA, IMA models are not regressive

[@FuZZbaLLbee]

If you add regressors to Prophet, doest that also make it multi variant?

[@yasserothman4023]

What about linear regression moving average and autoregression and Taylor series methods ? Why they weren't discussed

[@oxfordsculler8013]

Really good. Any chance of saying in the future how LTSMs and GRUs compare with the aforementioned models for time series.

[@RidingWithGerdas]

Seen many examples when LTSM wasn't much better but definetly was slower compared with a prophet for example. Mostly you need to spend more time doing data cleansing and feature engineering rather than searching models :)

[@youngzproduction7498]

Well done!

[@CodeEmporium]

Thanks!

[@muchostudios3716]

Great

[@aniruddhadatta925]

Well I know Reinforcement learning is being used to model financial time series control aka the stock market but can we track back and make a simple forecasting model with Reinforcement learning in which the actions dont really have nothing to do other than predictions

[@SP-db6sh]

Same thing I'm looking for !

[@ai95]

Danke je wel!

[@connectrRomania]

thank you

[@CodeEmporium]

Welcome

[@wallace2276]

Excellent summarization of the concepts in timeseries! Would love for you to make a video on the catboost algorithm and how to use it for timeseries forecasting problems.

[@CodeEmporium]

I'm thinking of Catboost for sure. Maybe a future video. Thanks for the suggestion :)

[@nikhildoye9671]

Hey, nice work. For the next video, can you implement a Temporal convolutional network for time series forecasting(Load Forecasting)?

[@CodeEmporium]

Sounds like a good idea. Maybe a future video. Thanks for the comment

[@rajnishadhikari9280]

Do you have project with multivariate analysis?

[@denvercabanes]

I'm trying predictive sequence or whatever it's called, from 0 knowledge in programming. Will prophet be a good starting point to get my feet wet?

[@CodeEmporium]

Would recommend starting Prophet if you want an easy start point for tone series analysis (time series forecasting from a few lines of code). However if you're started out with the field as a whole, I'd start with basic classification and regression problems. Maybe start with Linear Regression with a python library called scikit learn, then maybe ease into the math, repeat for different algorithms. Just keep pushing through regardless of how difficult it may seem i promise it'll become easier as you know more. The field will be fun too. Join us on Discord (link in the description of my latest vidoes). We can talk nore there :)

[@veloenoir1507]

Would it not be interesting to try and apply transformers to these kinds of problems? In the sense that price movement is a written text and we try and predict what is going to be said next.

[@aniruddhadatta925]

Or instead of modelling it like NLP we model something called convolutional Attention to deal with the raw time series float values data.

[@rajnidahiya6097]

Can we add categorical variables as explanatory variables as well or the variables should be time variant?

[@CodeEmporium]

You can def add categorical variables as features too

[@oraz.]

oh snap...

[@cipshadow]

Great video! What is it that makes Prophet NOT a machine learning model? I would have thought all the models that learn from the past can be classified as machine learning.

[@AbdulWahab-cy9ib]

Why can't we extend the machine learning model which predicts for next day to the 3days or 10days by simply using 1 feature which keeps track of the previous day? I can't see why can't we extend.

[@CodeEmporium]

Because for a 10 day out prediction, the "previous day" hasn't happened yet (at the time of making the prediction)

[@tactusxii]

How to build ARIMA models in Python without dates? If I'm estimating a target boats sinusoidal position in the ocean, do I wanna map milliseconds as dates 🤔, nah

[@lbers238]

Why shouldn't you be able to use Machine Learning recursively to predict more days into the future?

[@nmay231]

I'm not an expert; this is just my intuition. I think he says that traditional/general-case ML algorithms can not recursively predict more days into the future because they are generally not as good as time-series models (for time-series data at least). So. while both types of models will get progressively worse the more days you recursively predict values, ARIMA and similar models will "hold out" for longer. In essence, it's not impossible to recursively predict with traditional ML; it's simply less productive, especially since you can set them up to predict +3 days directly (not +1 +1 +1 day like in a recursive model).

[@adhoc3018]

I guess that errors would get amplified with each new call to the model.

[@danielhorn9691]

SARIMAX is multivariate

[@user-pl5zx4sp5c]

Why would my grandma has a laptop repair shop

[@Jacen_Rockwell]

I'm sure you're aware of certain psychological phenomena & "fringe theories" emerging on the surface? Some people, while not versed in the particulars, nontheless are beginning to trace the origins of said phenomena to, vectors adjacent, or even linked with our area of this continuum. I'm trying not to use any words that would attract scrutiny but that's becoming difficult. That narcissistic zealot, blabbing about timelines & Lovecraft. who's ass we are encouraged to kiss ROWS G.D QuÆv. I know people are getting the same threats I am...Either someone has worked out the Quasi-cosmic code, they've copied the artefact, or the אºD's are angry. Maybe we should focus on what's important. BTW I ASSUME NO-ONES SEEING DARK RECURRING NUMBERS & SEEING DEEP LEARNING OBSERVER NOISE? ME NEITHER!

[@jmoz]

You can make up any term and I wouldn’t know if it was real or fake. Dog apostle maxer.

[@MOUNICANUTAKKIPHD]

i need a help from you regarding time series ....how can i contact u

[@BENJAMINELEKWACHI]

please how can i contact you