วีดีโอ

The dropout rate is from the model specification. But yeah... actually we could try out some other values! I did play around with it briefly, but then totally forgot :D Thanks for your suggestions!

Hello, that's my plan, so stay tuned!

Glad you liked it! More to come :D

💪😄

Thanks for the suggestion. I'll see if I can make a video about that in an upcoming video 📚

Hermite cubic is actually a variation of cubic spline (and so does the B-spline). If you're talking about the natural cubic spline, then the short answer is no. The interpolation functions available in the library are only Nearest neighbor, Piecewise linear, and Cubic (through Hermite cubic). Theoretically, though, you might be able to use a custom interpolation function, although this could be more challenging, especially if we're talking about the natural cubic spline interpolation. That's because the cubic spline requires the first AND second derivatives at each node. In Hermite cubic, though, we only need the function value and its first derivative at each node. Please note that the answer I just gave you only results from my thinking process based on my understanding of cubic splines and Hermite cubic. The choice of a specific interpolation function is never mentioned in the papers, so I might be wrong. Hope this helps!

More to come! 😃

Thanks for the kind words!

😄

yeah, there are so many variations of DTW

@@TimelyTimeSeries keep going I need you I'm self learner I need more knowledge I'm seek in math but I use the models I believe that not the models predict the future but God see you working and the low of the life is the moving to something create the opportunity and bias

Thereis oppurtunity to make code together

@@IndraWiharjaa ok lets do it

Worth trying. Thanks for the insight!

I haven't tried applying Rocket since we don't have the implementation code for datasets with unequal length of inputs, unless we preprocess the input by padding. I didn't give any padding to the input (haven't tried it either... perhaps I'll try it in the future) because I want to try using minimum data preprocessing. So, I use MiniRocket in this video because the implementation code is available in sktime library. And yes, MiniRocket works better. Maybe Rocket could work as well because according to the paper, MiniRocket, theoretically, gives roughly the same accuracy as Rocket. The huge difference between MiniRocket and Rocket is that MiniRocket works much faster than Rocket.

Thank you I'll try my best!

Thanks for your kind words. Glad you enjoyed it!

Thank you, Professor Keogh!

Thank you so much!

Hi, thank you!

It might work. Recurrent Neural Network is one of several networks built for processing sequences.

Hi, thanks for your kind words. Yeah, sometimes I got those high variances. The variance naturally increases as we forecast further ahead, but if the variance is already too high right from the very next timestamp, I guess it means that our model is not confident of the forecast? So maybe decreasing the value of dW in dlmPoly function (or any other DLM model you use) could help since the system variance, W, is one of the components of Q--the variance outputted by the dlmForecast function. When working on this project, I found out (and supported by the textbook) that the W could affect our forecast performance greatly. Or maybe you should try another DLM component like 'dlmPoly' with order of 2 or 3. Let me know if you have further questions. Hope this helps!

@@TimelyTimeSeries ​Amazing. I'll try that. Thanks a lot. Just for curiosity: my models consist of a level component (dlmModPoly) and a seasonality one (dlmModTrig). Unlike your model, I do not put values for the variances (V and W) of the level component. Instead, I estimate them by likelihood (with the dlmMLE function) and the output values are quite low (on the order of 10^-1). Maybe this is the problem, right?

@@jaironleal4287 It could be, especially if the distribution is not Gaussian (even though it might be difficult to prove that the series follows a Gaussian distribution). Or if you have 'wrong' components. I tried using dlmMLE in one of my videos (in my second video to be exact), and my prediction from that model was not way better than this video's prediction. Btw, I might be wrong, but I have an idea. Maybe you can try to just multiply (or divide) both V and W by 10 or 100 if you want to change those values but keeping the ratio. It just came to my mind that perhaps the ratio of V and W outputted by dlmMLE is the good ratio, but the numbers itself should be increased/decreased.

@@TimelyTimeSeries ​ @TimelyTimeSeries Oh, I got it. In this case, I'd manipulate the 'signal-to-noise ratio' which means V/W. I'll do it. I'm sorry if I'm boring you but (if you allow me) I'd like to ask one last question (I promise 😅): have you ever needed to use a discount factor when using the dlm package? Maybe this could be a good feature, but I've never been able to understand if it's possible to do this with this package.

@@jaironleal4287 It's okay, don't worry 😀I really want to try using discount factor because, as you said, that really could help. But unfortunately, to my best knowledge, we can't apply discount factors with this package. And up till now, I haven't found a package in R or Python that could implement it.

Hi, thank you for your appreciation! I am still experimenting with audio, so thank you so much for your feedback. I will definitely try to reduce the echo in my future videos.

Unfortunately, no😅 The real values, shown at the end of the reel, are lower than my predictions. So, while the model's predictions say that the bond yield would keep increasing, the real values keep decreasing. However, the bond yield did increase later to 4.27%, which is the model's prediction, in the following month (September 2023).

​@@TimelyTimeSeriesI like your machine learning videos. I particularly liked the Bayesian video. Many TH-camrs show how they approximate the function of the security using the Lstm model and other models... but they have never predicted a price in the future using these methods, they use many methodes to predict the past price but not the Future price . I'm following a method that was developed by someone. This strategy is about deep forest algorithm, this method is about decision trees. With this method, the model learns to discover the same patterns that occur again and again, which are generated by the algorithm in the past in the form of triangles, squares, ovals in charts and copy them into the future chart. you can find these patterns yourself. The problem with this method is that you have to buy candlestick data for example closing price high low.. from Bloomberg terminal so that the program achieves more accurate results. Every TH-camr uses free databases like Yahoo. If you were to buy candle data using the methods shown, you would have high precision and high accuracy. I no longer follow the guy who developed this method but i use his method. He spend 30.000€ candle Stick data in month

Hi @metehan9185 , thanks for your feedback! And yeah, good data is really important so we can generate accurate predictions. Regarding deep forest, I see a lot of good time series forecasting models using deep forest (they didn't forecast bond price though). I plan to learn more about it soon.

@@TimelyTimeSeries Hi What I can tell you is that the person who developed this method, with the repeating patterns... has other mathematical calculations built into his program. He also incorporated Fourier transformation, because Fourier transformation is considered the most accurate iteration in machine learning, and he incorporated magnetic effect lines into some charts. These are lines that he calculated using Fourier transformation and said that the chart of a stock, for example Apple, must fall or rise to this line. He also uses a program to filter unnecessary price data such as knockouts. The program should be called Isolation forest. Using the decision trees, the program recognizes the same pattern that was filtered in the past with the Programm Isolation forest. He adds another important line, this line he calls main drive shafts. He calculates these main driving waves from the average value of the warrants to predict that the price could rise or fall to these lines. I think he uses mathematical models such as the Black Schols model, GARCH model, binomial model, Cox Ross Rubinstein model and averages these models to get to the main drive shafts. He thinks every strategy is bad and only works with this method. this method has a win rate of 84.4%. Too bad he had a TH-cam channel and deleted it. He has a TikTok account and streams there daily and explains his method every time. but only speaks German. But he also gives courses in English.

Hello, Mr. Keogh, nice to meet you. Thank you for saying that! I read your paper many times to understand Shapelets better. Your explanation of Shapelets by using leaves made me understand Shapelets immediately.

thanks for your kind words@@TimelyTimeSeries