I am a Professional Practice Assistant Professor in finance at Utah State University, Huntsman school of business. I recently moved to the Data Analytics and Information Systems department as I am spending more time on machine learning and deep learning concepts. Prior to joining the Huntsman School in 2018, I was a research associate within Financial Modeling Group at BlackRock NYC. My research fields are machine learning and deep learning applications in finance, empirical asset pricing and advanced time series forecasting.
In this channel I will be publishing my lectures in Econometrics, Machine Learning, Deep Learning and Advanced Time Series Forecasting (Deep forecasting)! I'll update the channel every week or two, and I'll answer as many questions as I can. You can find all the materials on my GitHub page: github.com/PJalgotrader
Machine Learning is different from general programming in this way: Machine Learning involves designing algorithms that can learn from data and improve over time, whereas general programming requires explicit instructions for every step. In Machine Learning, the focus is on enabling the system to recognize patterns and make predictions based on the data it processes. There are three main types of Machine Learning: 1. Supervised Learning: Both input and output are provided, meaning the labels are known. The algorithm learns to map inputs to outputs based on this labeled training data. 2. Unsupervised Learning: Only the input is given. The machine must detect meaningful patterns or structures within the data without any labeled outcomes to guide it. 3. Reinforcement Learning: Both input and output are provided but not at the same time. The system interacts with the environment, receiving feedback through trial and error, and learns to make decisions that maximize cumulative rewards. Classification of the following algorithms according to the types of Machine Learning: 1. Both input and output are given, meaning the labels are known. - Type: Supervised Learning 2. Only the input is given. The machine should detect meaningful patterns. - Type: Unsupervised Learning 3. Both input and output are given, but not at the same time. There might be a delay. The machine needs to explore the environment through trial and error until it discovers a meaningful pattern. - Type: Reinforcement Learning ai.walsoftcomputers.com/
In real-world problems where the emphasis is either on prediction or inference, the choice between Machine Learning (ML) and Statistical Learning (SL) can depend on the specific goals: Problem Emphasizing Prediction (Less Inference): - Approach: Machine Learning (ML) - Reason: ML is generally more suited for problems where the primary goal is to make accurate predictions or classifications based on large datasets, often without needing to understand the underlying relationships between variables in detail. ML models like neural networks, ensemble methods, and gradient boosting are designed to optimize predictive performance and can handle complex, high-dimensional data. Problem Emphasizing Inference (Less Prediction): - Approach: Statistical Learning (SL) - Reason: SL focuses on understanding and interpreting the relationships between variables, often through simpler, more interpretable models. Techniques like linear regression, logistic regression, and hypothesis testing are used to infer relationships and test hypotheses about data. These methods are designed to provide insights into the underlying processes and can be used to draw conclusions about causal relationships or other aspects of the data structure. In summary: - Use ML for problems where accurate prediction is crucial and interpretability is less of a concern. - Use SL for problems where understanding relationships and making inferences is more important than achieving the highest predictive accuracy. ai.walsoftcomputers.com/
Pls can i use the vprom autoencoder with mcmc model or hmc model to catch daily & weekly forcast im trying use combination of economic reports and daily financial data im self learner i think its kind of art mix or im wrong ?
Hi, Combining complex models like autoencoders with MCMC or HMC for predicting daily and weekly stock prices sounds interesting, but none of these methods have been proven to consistently work for short-term financial forecasting. The stock market is very unpredictable and full of random noise, which even advanced models struggle to handle well. As a retail investor, I highly encourage you not to waste time trying to beat the market with such models. The chances of success are low, and it’s better to focus on long-term, proven investment strategies. Good luck, and feel free to ask if you have more questions!
@@pedramjahangiry yes you absolutely right the market is hard to predict I'm trying put my knowledge about data and the new release in my ideas 💡 It is only a quest for sustenance. I am a Muslim and Islam denies the idea that anyone can predict anything because sustenance comes from God.
When using timeseries_dataset_from_array to create datasets, train_dataset, test_dataset and val_dataset had all uniform tensors except the last one which were partials, i.e. their samples and targets were as follows: for samples, targets in train_dataset: if samples.shape != (batch_size, sequence_length, 14): print(samples.shape) print(targets.shape) for samples, targets in test_dataset: if samples.shape != (batch_size, sequence_length, 14): print(samples.shape) print(targets.shape) for samples, targets in val_dataset: if samples.shape != (batch_size, sequence_length, 14): print(samples.shape) print(targets.shape) (103, 120, 14) (103,) (118, 120, 14) (118,) (206, 120, 14) (206,) This gave the error: Epoch 1/10 --------------------------------------------------------------------------- InvalidArgumentError Traceback (most recent call last) <ipython-input-146-e09ecdf9d4ec> in <cell line: 15>() 13 ] 14 model.compile(optimizer="rmsprop", loss="mse", metrics=["mae"]) ---> 15 history = model.fit(train_dataset, 16 epochs=10, 17 validation_data=val_dataset, Only one input size may be -1, not both 0 and 1 [[{{node functional_9_1/flatten_10_1/Reshape}}]] [Op:__inference_one_step_on_iterator_41264] Although I remove the partial batches, I still get the error. I do not get the same error when fitting the dataset with the CNN.
I discovered your videos yesterday, and I have to admit that they are exceptionally good.I have long been wanting to advance my knowledge and skills within time series forecasting but haven't been able to find a course that seemed worth my time. However, I have been binging the Deep Forecasting course - it just perfectly matches my current background knowledge and skills within programming, it covers the methods and models that I am interested in mastering, and everything is explained really well. I cannot wait for the coming modules!
of course it depend on the dimension of features and patterns in data. but, "generaly speaking", I would rank them like this: Xgboost> random forest> SVR > KNN
oh my dear god, it is the greatest ML videos I have ever seen in my life, I cant understand the concepts or I get bored, but I can watch this videos everyday all day long, thanks dear Pedram <3
The concepts are so well explained. I love the way each phase of the process is taught in a clear way. Im better able to grasp concepts of machine learning now with clarity. Thanks for you efforts!
I would highly encourage to learn them side by side. It may slow you down but it is worth your investment. It is good to start with fundamentals of regression analysis. After all, machine learning is the extension of statistical learning I believe. Good luck!
This is my personal experience (not so good one) with the PyCaret library. *First Disappointment:* After installing it with `pip install pycaret` on my Linux machine, I started Python from the command line and ran `import pycaret`. This is the message I received: ``` RuntimeError: ('PyCaret only supports Python 3.9, 3.10, 3.11. Your actual Python version: ', sys.version_info(major=3, minor=12, micro=4, releaselevel='final', serial=0), 'Please DOWNGRADE your Python version.') ``` Alright, I can live with that. After creating a virtual environment with a downgraded Python version and playing with PyCaret for some time, I was quite impressed by the rich capabilities-until I tried to analyze the performance of a trained model. *Second Disappointment:* My script contains the following line: ```python experiment1.plot_model(tuned_gbr, plot='learning') # Learning Curve ``` This line does work and I get to see the plot, but when I close the plot by clicking on the "X" mark or pressing Ctrl+W, the script crashes with the following output: ``` File "/usr/lib64/python3.11/tkinter/__init__.py", line 1732, in __setitem__ self.configure({key: value}) File "/usr/lib64/python3.11/tkinter/__init__.py", line 1721, in configure return self._configure('configure', cnf, kw) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib64/python3.11/tkinter/__init__.py", line 1711, in _configure self.tk.call(_flatten((self._w, cmd)) + self._options(cnf)) _tkinter.TclError: invalid command name ".!navigationtoolbar2tk.!button2" ``` Actually, any plot type will crash in the same way except for the 'pipeline' plot, which does continue executing the remainder of the script after closing. This is very unfortunate, as I was looking forward to the amazing PyCaret plotting and analysis capabilities, which are now completely useless as they crash the program. Maybe it all works in Jupyter Notebook and Google Colab, but I don't use those platforms. If you can't run a simple script, it ends there for me. *Conclusion:* PyCaret is an excellent resource for learning. I'll try to adopt its concepts, pipelines, and methodologies, but I'll have to write my code myself as this thing unfortunately doesn't work. I hope developers will continue improving it and maybe it will become better with time. For now, it's a learning material, which I'll definitely examine in more detail. Just by going through the PyCaret documentation, I am fascinated by how much there is to learn.
Hi Pedram, I really need your help. I'm working on a project that detects seasonality using pycaret setup and it's all working great except for that the setup function shows seasonality present even with the most randomest data eg:- just a data of random values ranging from 20-25. Is there a way to fix this?
The pycaret setup might sometimes detect false seasonality in random data. To fix this, try manually inspecting your data and using statistical tests to verify seasonality. You can also adjust the setup parameters or set seasonality=False manually. Hope this helps!
And here we are new video from the awesome 😎 teacher ❤ i need to build a trading ideas from sataliet data pls go to the moon with this course 🙏🏻 thank you
I am trying to learn ML in Azure and Gemini and all I keep being told is to use different models. What I am interested in is understanding how models are build and the though process that goes into doing it. Your course answers this question really well. Thank you.
it is in the deep learning course. Find it here: github.com/PJalgotrader/Deep_Learning-USU/tree/main/Lectures%20and%20codes/Module%205-%20Deep%20Computer%20Vision/CNN_python
Hi Pedram, I was curious while watching. Let's assume that I make a model that is very complex, I run through a large number of samples, and I end up with very high variance. Why would I not be able to store all these individual parameter sets from each run on the sample sets, average of them and then find a model that fits this averaged line which I could then use as my true model? For my mind, it seems like this average finds the true relationship quite well so would avoid the overfitting issues of making overcomplex models and could possibly be a better fit then the balance model. Also, I just want to say whether you are able to respond or not, this is a fantastic series. I had found it with videos 11/12 in the process of trying to understand some SciKit-Learn docs for a project and once finished with those I went back to video 1 and am now going through everything. I plan to go to the DL and ML in Finance playlists after and it is truly outstanding that you put these up for free on youtube, it is exactly what I love about the internet and I want to sincerely thank you for being a part of it. It has somewhat distracted me from my project directly but the information is wonderful and you are an excellent teacher so it is a worthy distraction!
Hi Pedram, one question. Why do you scale the whole dataframe in this notebook? In the other notebooks you only scale the X's; in particular, you define X_train_sc = sc.fit_transform(X_train) and X_test_sc = sc.transform(X_test). However, in this notebook you do df_sc= scaler.fit_transform(df), where df contains the y and the Xs. So in this case, you scale both (y and X) and potentially introduce some bias by fit_transforming all: the train and the test split (since you take them from df). Why doing that?
Good point! Yes, we typically use the training set for fitting the scaler and then apply it to the test set. However, for this dataset, as I mentioned in the video, I was a bit lazy and scaled the entire data since it's a small cross-sectional dataset with no outliers. You can certainly do it properly (fit_transform(X_train), transform(X_test)) and let me know if you find any differences. Regarding scaling 𝑦, it doesn't impact model performance much but can help with numerical stability and faster convergence during optimization. The model's predictive power remains the same. Interpreting RMSE would be different, though, because the unit is scaled.
Excellent lessons. But I have one problem though. Wouldn't the variance of the ensembled prediction be the sum of all var of singles trees( and the cov term) divided by the square of B? Or am I getting something wrong?
