Nathan Burton

HIGH-FREQUENCY CAUSAL MODEL OF THE U.S. ECONOMY STOIC is currently developing a high-frequency causal model of the U.S. economy using the 490 Brave-Butters-Kelley monthly indicators (BBKI). Future developments might include daily nowcasting series. The discovery of causal relationships between BBKI indicators is handled in two complementary manners: first, a Large Language Model (GPT 4o) is used to suggest possible causal relationships; second, statistical mining techniques inspired by Sadeghi, Gopal, and Fesanghary (2023) are used to discover relationships unknown to the LLM. Causal relationships are then formally established using do-calculus. The persistence of causal relationships will be tested across the entire period covered by the indicators (1960 to present). Regime-dependent relationships will be investigated in relation to periods of recession, as defined by the National Bureau of Economic Research (NBER). This causal model will be outlined in an Excel spreadsheet and implemented as a Python library directly available from Excel using the STOIC Jupyter add-in for Excel (Cf. screenshot). Interactive causality graphs powered by Quarto will be provided as well. STOIC is currently discussing with various data providers to provide daily nowcasting versions of the 490 indicators. Such higher-frequency indicators should make it possible to detect (not just predict) a recession at least one month in advance, ahead of BBKI, CFNAI, or Sahm Rule. Please contact us if this work is of interest and you would like to join our early adopter program. Also, if you happen to work for a Central Bank or Sovereign Fund, you might be interested in the development of such a causal model for your domestic economy.

29

Quantpedia in October 2024 – An enhanced Basic Overview report – new calculations and reporting capabilities – 11 new Quantpedia Premium strategies have been added to our database – 9 new related research papers have been included in existing Premium strategies during the last month – Additionally, we have produced 7 new backtests written in QuantConnect code – 5 new blog posts that you may find interesting have been published on our Quantpedia blog in the previous month https://lnkd.in/dXmvigrV

2

The final Linear Algebra class of my "ML Foundations" curriculum (on Eigenvectors) is next week. After that, Calculus classes begin: Those run from May to July and registration is open for most Calculus classes now. Hope to see you in class! I've been impressed by the attentiveness, curiosity and plain old intelligence exhibited by the students in the Linear Algebra classes, several hundred of whom have shown up to every class so far. My "Machine Learning Foundations" curriculum is tailored to provide you with the foundational knowledge you need to deeply understand all contemporary ML approaches, including deep learning, Large Language Models (LLMs) and A.I. in general. The content is brought to life through a combination of: • Vivid full-color illustrations • Paper-and-pencil comprehension exercises • Hundreds of Python code examples within hands-on Jupyter notebooks (with a focus on the PyTorch and TensorFlow libraries) • Practical ML applications • Resources for digging deeper into topics of interest And, as highlighted by the graphic on today's post, the content is organized into four subject areas: • Linear Algebra for ML • Calculus for ML • Probability Theory and Statistics for ML • Computer Science (Algorithms, Data Structures and Optimization) for ML Within a subject area, the content builds (e.g., Calculus III depends on knowledge of material in Calculus I and Calculus II), but the subject areas are relatively distinct (e.g., you could jump into the Probability subject area without attending LinAlg or Calc). Classes always begin at noon ET / 9am PT and run for 3-4 hours. There are a total of 14 classes in the series, with a class every 2-3 weeks. The first two classes (both on LinAlg) are behind us; here are the dates of the upcoming sessions with confirmed dates: 3. Linear Algebra III: Eigenvectors on May 8th* 4. Calculus Intro on May 22nd* 5. Calculus II: Automatic Differentiation on June 5th* 6. Calculus III: Partial Derivatives on June 26* 7. Calculus IV: Gradients & Integrals on July 10th 8. Intro to Probability on July 25th *Registration is open in O'Reilly now for these classes. Search for Live Events by author "Jon Krohn". If you missed sessions or can't wait for upcoming ones, use a web search for "github jon krohn ml foundations" to find my code repo and links to recorded content: • O'Reilly videos: All of the content is already available as a 28-hour video series. • YouTube: All of the Linear Algebra and Calculus content is in 100% free YouTube playlists. • Udemy: Contains all of the YouTube content, plus fully-worked solutions to exercises. For those of you who have been waiting for ages (over a year!) for updates to the YouTube/Udemy content, I'm teaching this series specifically to get me back into the swing of the material and get back to publishing. #machinelearning #ai #mathematics #calculus #livetraining #oreilly

137

11 Comments

Performance investigation of osmotically assisted reverse osmosis using explainable machine learning models: A comparative study https://lnkd.in/gPtPUBah •OARO performance was investigated under various operating conditions. •Ensemble ML models outperformed conventional ML models in predicting OARO outputs. •ML model performance significantly increased using the FO/RO membrane datasets. •SHAP analysis results were consistent regardless of the ML model used. •SHAP analysis revealed the latent contributions of some input variables.

21

This “feuding” is not doing Musk any favors… https://lnkd.in/gST4Mqcy I think Yann LeCun prediction on AGI (ie., it is not imminent) is correct, there’s just a lot of unsolved research and development problems to be done. Comparing that to Musk saying it’ll be next year, just indicates he doesn’t understand the issues with AI IMO. I think Musk’s track record in business predictions is just terrible. See… https://lnkd.in/gXSjuYX2 Plus, I’m not convinced AGI is certainly safe, or even a good thing for the socio-economy, in particular if it’s rushed to operation without adequate alignment (something Musk thought hitherto, but now has sidelined). I have concerns that AGI will be used to massively reduce human employment in the short term (eg., I’m skeptical about the practicality of UBI for potentially billions and that impact). Musk is pro UBI but has its impact been seriously modelled? Not really! See… https://lnkd.in/g7CeCQ2a I find his arguments with OpenAI and his response in investments in XAI inconsistent with his asserted concerns over AGI risk, simply not logically connected IMO. See.. https://lnkd.in/gc6PnKq4. And… https://lnkd.in/gCi5ZQFM Plus some of his companies seem to have troubles (eg., X and Tesla). See… https://lnkd.in/gXP7ABrt And… https://lnkd.in/g5PHjA95 Yet he is pushing for a 55 billion pay package! See… https://lnkd.in/gCyHehQk

18

3 Comments

Finished chapter 8 on amortized analysis There wasn't much covered in this chapter just 3 techniques like two pointers, nearest smaller elements (similar to next greater element question on Leetcode), and sliding window The main idea behind amortized analysis (or at least what I got from it) is to find the average time complexity This is because if you have 1 calculation that takes a lot longer then all of the other calculations during the execution of that algorithm, then instead of giving a worst case time complexity with the amount of time that 1 calculation took, we want to average out all the costly calculations I found this article quite useful for understanding it, https://lnkd.in/dtABkEzQ I only have 2 more chapters to go on range queries and bit manipulation before the first section of the book is finished!

3

Time series forecasting using shapelets and transformers, an excellent paper by Xuan May and her team. Definitely worth a read if you are interested in forecasting for finance. https://lnkd.in/gjGpx3Xy. #forecasting #ai

38

3 Comments

What is the probability that two positive integers x and y chosen uniformly at random will be relatively prime, i.e., gcd(x, y) = 1 (gcd(x, y) is the greatest common divisor of x and y)?

3

GEOMETRIC SHARPE RATIO If you like log returns and want to compute a Sharpe Ratio for them, please use this brand new definition of a Geometric Sharpe Ratio. Computing an arithmetic Sharpe Ratio on log returns simply does not make any sense. Please don't do that. Full construction available on this paper: https://lnkd.in/grxBec4W

209

96 Comments

My first link this week is about how to create charts at scale (https://lnkd.in/gSHq4N8s). In other quant guidance, here’s a piece on the %notin% operator in R (https://lnkd.in/gqmZnTyk), and some options for creating interactive tables (https://lnkd.in/gVxDumRQ). And once you’ve done your analysis, here’s some tips for how to confidently present to an audience (https://lnkd.in/g4uwQrux). Great news: bottlenose dolphins smile when they play (https://lnkd.in/gZd9szrD), and at least one species of bee is thriving (https://lnkd.in/gjAcJJWw). In maths news that I have a quite tenuous grasp on, a well-known conjecture in graph theory has been disproved (https://lnkd.in/gsXrdhmQ), and elsewhere I have come across an ingenious way to collect atmospheric data by using smartphones (https://lnkd.in/gtNFcC5d). Moving beyond the ionosphere, this astrophotography project beautifully straddles science and art (https://lnkd.in/g3nzHrq8), just as these sculptures straddle human and animal creativity (https://lnkd.in/gf-KjVRT). I also really dig these scratchy architectural illustrations (https://lnkd.in/gxhXG5hk) – perhaps one day I’ll be able to take night trains around Europe (https://lnkd.in/gbhG8gxs) and enjoy the architecture for myself.

1

Great stuff Armand!! This fits perfectly in my 3-pointers philosophy to effective R&D: 1. Simplicity over complexity. 2. Rapid experimentation to facilitate fast iterations. 3. Progression over perfection. Perfection through iterations. Neglecting any of these will challenge successful completion and delivery of any SW product whether #AI-based or not.

1

🎉 SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion (in PyTorch) Latest update to paper on ArXiv: 12-Jun-2024 👉 Article on Medium: https://lnkd.in/gnF2jJwP 👉 Paper: https://lnkd.in/gpEykKx6 👉 Code: https://lnkd.in/gr7SY32Z Paper Abstract: Multivariate time series forecasting plays a crucial role in various fields such as finance, traffic management, energy, and healthcare. Recent studies have highlighted the advantages of channel independence to resist distribution drift but neglect channel correlations, limiting further enhancements. Several methods utilize mechanisms like attention or mixer to address this by capturing channel correlations, but they either introduce excessive complexity or rely too heavily on the correlation to achieve satisfactory results under distribution drifts, particularly with a large number of channels. Addressing this gap, this paper presents an efficient MLP-based model, the Series-cOre Fused Time Series forecaster (SOFTS), which incorporates a novel STar Aggregate-Redistribute (STAR) module. Unlike traditional approaches that manage channel interactions through distributed structures, e.g., attention, STAR employs a centralized strategy to improve efficiency and reduce reliance on the quality of each channel. It aggregates all series to form a global core representation, which is then dispatched and fused with individual series representations to facilitate channel interactions effectively.SOFTS achieves superior performance over existing state-of-the-art methods with only linear complexity. The broad applicability of the STAR module across different forecasting models is also demonstrated empirically. #timeseries #python #deeplearning #ml #SOFTS #paper #code #pytorch

9

1 Comment

🚀 Exciting News! 📘 Our latest survey paper is now released, presenting a comprehensive analysis of hallucination phenomena in multimodal large language models (MLLMs), also known as Large Vision-Language Models (LVLMs). 🧐 Despite their groundbreaking capabilities in handling multimodal tasks, MLLMs face significant challenges, particularly the issue of hallucination—where models generate outputs inconsistent with the visual content. This not only hampers their practical deployment but also raises concerns about their reliability in real-world applications. 🔍 In our survey, we delve into the roots of this issue, reviewing the latest advancements in detecting, evaluating, and mitigating such inaccuracies. We explore a variety of causes, benchmarks, metrics, and strategies that are being developed to address these challenges. 🌐 Our goal? To deepen the understanding of hallucinations in MLLMs and to foster further advancements that enhance their robustness and reliability. This paper is a must-read for both researchers and practitioners in the field, providing them with valuable insights and resources. 🔗 Find all the details and more resources on our GitHub: Awesome-MLLM-Hallucination. The paper is located at: https://lnkd.in/gFf6XdDc 📢 Feel free to share your thoughts, questions, or insights on this topic! Let's push the boundaries of what MLLMs can achieve together! #MLLM #AI #MachineLearning #DataScience #Research #Innovation

137

2 Comments

You may have heard RNN/LSTM and the time-series LLM Lag-Llama. You may be curious about the development of time series modeling from RNN/LSTM to Temporal Fusion Transformers and then Lag-Llama. Besides showing you how to apply the techniques, this book explain the evolution. I'd like to make this post, which was behind the medium.com paywall, open to you. It is Chapter 19 in the book "Modern Time Series Forecasting Techniques." Have you pick up a copy? The ebook is made available on a secure platform teachable.com for readers who prefer the digital edition. Sample eBook chapters: https://lnkd.in/e9JX-qKa eBook on Teachable: https://lnkd.in/eRsFBsYT Amazon print book: https://a.co/d/25FVsMx https://lnkd.in/eSEgnhF9

10

This is why the research industry needs to take Synthetic Data seriously. https://lnkd.in/e65Jrs8S We need to see what works, what do we mean by works, how can we check it, and what guardrails do we need. The question is not whether clients start using Synthetic Data to answer questions, that has started. The question we need to answer is what our role will be. I disagree with Mark Ritson on several points in this post, not least the quality of the testing of Synthetic Data that has happened so far. But my main disagreement is when he says "It’s bad news for market research companies because they are likely to experience a steam-train decade of oncoming paranoia, pressure and partial obsolescence. Naïve marketers might expect these firms to adopt and adapt to synthetic data but, of course, the skills and competences required to talk to humans and program computers are entirely different and probably prohibitive." More than 50% of market research is already ResTech bases, the skills of asking the right questions are readily applied to AI situations, and to understand consumers, you have to understand consumers. Perhaps it is people offering mini-MBAs who will miss out ;-) Thoughts? Jon Puleston, Judith Passingham, Joaquim Bretcha, Simon Chadwick, Karine Pepin, Alexandra Kuzmina, Luisa Gibbons, Barry Jennings, Peter Harris, Xabier Palacio, Claudio Gennaro, Elaine Rodrigo, Nancy Deng, Sue York, James Ruta, Lauren Thompson

68

68 Comments