I think the core of the problem is that 1) past returns are just lousy predictors for future returns in shorter time frames, so garbage in garbage out. 2) A week worth of out-of-sample test is probably pure noise. FYI, time series momentum does exist, just on a longer time frame, check out this paper: Time Series Momentum by Tobias J. Moskowitz.

I know its beyond the point of the video, but I think its really important to mention friction/transaction/slippage/taker/maker cost of any strategy, can quickly turn positive expectancy negative and erode edge quickly.

This is great! I think, as you mentioned, that taking the actual price of the stock as the response variable in the LSTM could have made an important difference, and also make the results more comparable since then both the ARMA and LSTM would be modeling a continuous target. Can't wait to perform this exercise on my own and see what turns out. Thanks a lot for the great content!

Hi @ritvikmath, I recently came across your playlist on Time Series Analysis, and I found it incredibly insightful-thank you for that! I was wondering if you could make an additional video on ARIMAX, particularly focusing on how to incorporate exogenous predictor variables. That would be fantastic. Thanks again, and greetings from Germany!

This is a great series, I have been playing with that for some time and I know how hard this is. I have a suggestion, when you create a model you are not doing any backtesting on unseen data. You are just applying it to the next week, this is nice as entertainment but would be much better to have some backtesting stats to compare the unseen week. Would be nice to do a backtest on the last 4 weeks create some average and them apply it to the next week.

Excellent video! I'd like to see you try out this using reinforcement learning. It seems pretty easy to adapt your original LSTM model to something like a deep Q network or PPO. Love your stuff, please keep making more great vids. :)

Do you consider commissions?

good strategy kip up on this kind of experiment 👍

Great video as always! Have you ever tried modeling the Numerai data? It is a hedgefund, that released a a huge obfuscated financial dataset that is publicly available to model the returns of tickers in weekly timeperiods. Then it trades based on the predictions submitted by data scientists all over the world. It is more a data science challenge than financial modeling, but I think it would make an interesting video and I would love to hear your opinion on the project!

It would be interesting to see how the models perform on cryptocurrencies rather than stocks as they have fewer variables affecting their movement (no CEO, no profits made public, no product failures, etc). I have found that they strongly correlate to the typical market indices such as the sp500. Along with that, I am interested to introduce features such as market and stock volatility scores, maybe the previous days' sentiment analysis of news headlines pertaining to the stock/crypto, and the buy/sell/hold ratings of various analysts. also, are you predicting the % change as a regression problem, or predicting classes like Strong Decrease, decrease, neutral, increase, and strong increase as a class problem? I am working on a sentiment analysis of news headlines for cryptos dataset if you are interested!

I wanted to say that if we are doing day trade I think taking into account the day of the week as a feature might be a good idea.

thx for your... honesty! btw if you want t compare to a s&p500 on a weekly basis, you have to use models based on... weekly data! anyway, we know the results

It's will be very helpful, if backtest and training a model Live, that will be educational and informative, and thankyou ritvik sir for testing and full overview

It's maybe also important to take the vix and dxy to be sure the movement in the stock are not just related to that, I mean it's important to know the move is specific to the stock. Also the us03y - us01my can help but it's it's usually too volatile because it's already a % and also some time by just taking variations we lost the sense of the value and some kind of thresholds.

do you think 5 days is sufficient to attain any statistical confidence? how many days would we need to run the experiment to have a certain threshold of confidence?

Very interesting series, thank you! I have a funky suggestion.. How about a bit of crypto market madness? :)

One model outperforming another could just be due to chance, right?

I was just wondering, is it reliable to use the model for trading? Because I don’t have any experience on picking stocks and doing trading. But I want to do it. What would be your opinion on using these kind models for day trading? Also, do we also need to have some strategy to trade using the algorithms?

By now I'm sure you've got a super-extra-amazing-deluxe model for your private investments🧠

If AI really works for stocks, all the ML engineers will be working on this.

Lstm is a black box. Prone to overtraining. Use models as simple as possible. Old school statistical models are the best.

you know what the sad reality is. Since interest rates are so high. if you get high yield savings of 5.1 APY. It's way safer than trading and you won't get draw downs. LOLs. i watched all this ML videos and python videos. Did all these sleepless nights in trading. Even with any trading strategy it's still 60%. You should back tested by making trade with trading view manually with your hands. Then coding it. Simulated trades do not equal real trades. Just buy and hold for 30 years it will 200% the Top 10 SP500 stocks.

how about using the transformer model?

Use your statistical skills for something more valuable is my advice

Super