Jeremy Kulcsar

Our latest preprint on Adaptive ML is now available on arxiv 💥 ! In this paper entitled “Evaluating the Efficacy of Instance Incremental vs. Batch Learning in Delayed Label Environments: An Empirical Study on Tabular Data Streaming for Fraud Detection”, we question the superiority of instance incremental algorithms (in terms of predictive performance/interpretability) when we take into account real conditions such as label or feedback delay, class imbalance 😑.  Our results using real world fraud data (as well as generated data widely used in academia) clearly invite researchers to rethink the design and evaluation of supervised incremental machine learning algorithms, with the aim of narrowing the gap between pure academic research and industrial applications (for example, in fraud detection). link: https://lnkd.in/eJdQYgQM code: https://lnkd.in/eWrRvHP9 Thanks to my PhD advisors and co-authors Mustapha Lebbah, Hanene Azzag, Grégoire Jaffre and Zaineb Chelly Dagdia. #FraudDetection, #DataStreams, #IncrementalLearning

44

Forget about LLMs Ranking Leaderboards🥇, Welcome to the Era of Model Routing 🔀! This shift in the market is truly transformative. The idea of dynamically selecting the most suitable model for each specific task is a game-changer. It's like having a garage full of specialized vehicles, each ready to take on a different terrain. Whether it's the speed of a Ferrari 🏎 for complex technical analyses or the reliability of a Range Rover 🚗 for varied tasks, model routing ensures the optimal performance for every scenario. Amazon Q for Business is a perfect example, dynamically choosing the best model based on query complexity to maximize user experience. This approach heralds a future where model selection is flexible, task-specific, and efficient, marking a significant departure from static model use. Exciting times ahead in the world of generative AI! #AI #MachineLearning #ModelRouting #Innovation #TechRevolution #FutureOfAI #GenerativeAI #SmartTech 🚀🔍🧠

1

🤖 MetaGPT Redefines Multi-Agent Collaboration 🤖 📄 arXiv 𝗣𝗮𝗽𝗲𝗿: "MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework" by Sirui Hong et al. During my talks, people usually ask me where I expect to see a boom in the incoming months. If Sam doesn't find AGI in the short term 😂, my answer is clear: Agents, Agents everywhere! So let's take a look to this paper about Agent Collaboration. 📝 𝗔𝗯𝘀𝘁𝗿𝗮𝗰𝘁 𝗦𝘂𝗺𝗺𝗮𝗿𝘆 Take a look to MetaGPT, a game-changer in multi-agent collaboration. By leveraging Meta Programming techniques, MetaGPT optimizes collaboration efficiency through Agent Roles, Workflows, Structured Messages, and Efficient Communication Protocols. Its outstanding performance on benchmarks like HumanEval and MBPP sets a new standard, particularly excelling in software development tasks. 𝗧𝗟;𝗗𝗥 🧠 MetaGPT introduces cutting-edge Meta Programming techniques for enhanced multi-agent collaboration. 📊 Achieves unparalleled performance on industry benchmarks like HumanEval and MBPP. 👨‍💻 Excels in software development tasks with high task completion rates. 🤝 Incorporates structured messages and efficient communication protocols for streamlined collaboration. 👾 Ablation study underscores the significance of diverse agent roles in boosting performance. 𝗙𝗼𝗿 𝗺𝗼𝗿𝗲 𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 👇 📜 ArXiv Paper: https://lnkd.in/dt7XuM3Z 📪 Andrew Ng post on Multi-Agent Framework: https://lnkd.in/d5pmYcpB 👩‍💻 MetaGPT Github repository: https://lnkd.in/dQa7T4wM 👏 Kudos to Val Andrei Fajardo for the amazing visualization of the pager, it's pure magic for the eyes!

53

4 Comments

[𝗟𝗟𝗠-𝗧𝗶𝗺𝗲𝗦𝗲𝗿𝗶𝗲𝘀] CALF: LLM alignment for #timeseries forecasting through fine-tuning between different modes of transport 𝗝𝗼𝘂𝗿𝗻𝗮𝗹 𝘁𝗶𝘁𝗹𝗲 𝗮𝗻𝗱 𝘆𝗲𝗮𝗿 𝗼𝗳 𝗽𝘂𝗯𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻: arXiv, 2023 𝗙𝗶𝗿𝘀𝘁 𝗮𝗻𝗱 𝗹𝗮𝘀𝘁 𝗮𝘂𝘁𝗵𝗼𝗿𝘀: Peiyuan Liu, Hank0626 𝗙𝗶𝗿𝘀𝘁 𝗔𝗳𝗳𝗶𝗹𝗶𝗮𝘁𝗲𝘀: School of Computer Science, University of Science and Technology of China __________________________________________ 𝗔𝗯𝘀𝘁𝗿𝗮𝗰𝘁: CALF is a new cross-modal fine-tuning framework designed to align large linguistic models (LLMs) with time series forecasting tasks. By addressing the distribution mismatch between text and time data, CALF leverages the robust capabilities of LLM to increase the accuracy and efficiency of multivariate time series (MTSF) forecasting. The framework introduces key components such as Cross-Modal Match Module, Feature Regularization Loss, and Output Consistency Loss to ensure efficient modality matching and consistent semantic context. 𝗕𝗮𝗰𝗸𝗴𝗿𝗼𝘂𝗻𝗱: A problem in applying LLM to time series forecasting stems from the significant modal mismatch between text and time data. Traditional methods using LLM directly with temporal data often lead to suboptimal performance due to this mismatch. 𝗠𝗲𝘁𝗵𝗼𝗱𝘀: Cross-Modal Match Module: Integrates time series and text inputs using keyword embedding extraction and mutual attention mechanisms. Feature Regularization Loss: Aligns the intermediate layer outputs to efficiently guide gradient updates. Loss of output consistency: Ensures that the output representations of text and time series modalities are consistent. 𝗥𝗲𝘀𝘂𝗹𝘁: The CALF framework has demonstrated improved performance in MTSF tasks by effectively aligning modalities and leveraging LLM strengths. Metrics such as MSE (mean squared error) and MAE (mean absolute error) were used to evaluate model performance across different datasets. 𝗗𝗶𝘀𝗰𝘂𝘀𝘀𝗶𝗼𝗻: The CALF approach significantly reduces the distribution mismatch between text and time data, thus improving the ability of LLM to handle time series forecasting tasks. The framework's crossmodal techniques ensure that the learned representations are efficient and effective, leading to better forecasting performance. 𝗡𝗲𝘄 𝗰𝗼𝗺𝗽𝗮𝗿𝗲𝗱 𝘁𝗼 𝗽𝗿𝗲𝘃𝗶𝗼𝘂𝘀 𝘀𝘁𝘂𝗱𝗶𝗲𝘀: CALF introduces a cross-modal fine-tuning strategy that bridges the gap between textual and temporal data, a novel approach compared to traditional methods that typically do not address this modal mismatch. 𝗟𝗶𝗺𝗶𝘁𝗮𝘁𝗶𝗼𝗻𝘀: The main limitations are the computational cost associated with cross-attention mechanisms and the need for dimensionality reduction techniques such as PCA, which may not be efficient for extremely large datasets. 𝗣𝗼𝘁𝗲𝗻𝘁𝗶𝗮𝗹 𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀: The CALF framework uses large language models to improve the accuracy and efficiency of time series forecasting, opening many possibilities in predictive analytics.

77

4 Comments

It's fascinating to see how far AI has come in processing complex financial data! 🏦📊 Claude 3.5 Sonnet's ability to accurately extract detailed information from intricate charts is a testament to the rapid advancements in AI. This precision will undoubtedly revolutionize financial analysis and reporting. I'm curious about the level of precision Claude 3.5 Opus might achieve, given the significant improvements in Claude 3.5 Sonnet. These developments are exciting and hold great potential for enhancing our analytical capabilities in finance and beyond! 🚀🔍 #AI #MachineLearning #FinancialAnalysis #DataScience #Claude3.5 #TechInnovation #ArtificialIntelligence #FinancialReports #Innovation

3

[Free Python Workshop] Algorithmic Trading: From Trade Strategy to Responsible Execution This is what we're teaching live: Learn how to develop trading strategies, backtest them, and execute them all in Python. Inside the LIVE EVENT, we will demystify: 1. Which algorithmic trading strategies to avoid (and which strategies hedge funds actually use) 2. The secret to turning 1 foundational trading strategy into 100 testable trading strategies (Never run out of trading strategies again) 3. Why traders fail with "profitable strategies" (the secret to trading responsibly) 4. The Python Quant Stack: Which Python libraries we use to get 90% of our results. 5. How we automate trade execution (trade automatically just like hedge funds) BONUS Live Python Demo: How to trade a volatility targeting strategy using Python This is a Python Coding Workshop, so Python experience is a plus. What's your next step? Register here (500 seats): https://lnkd.in/eMKQPTWs

19

1 Comment

RECONSTRUCTING HISTORICAL TRADING PnL: A DATA PIPELINE APPROACH Bien Vo (our engineer) outlines a robust solution for missing historical trading profit and loss (PnL) data. Gaps in records lead to: - Traders: Unable to validate past strategies or measure performance. - Platforms: Struggling with compliance and reporting. 1/ The Challenge: Missing historical records - High Volume: 43,200 data points per token per month. - API Limits: Binance caps requests to 1,500 points each. - Complexity: 300+ positions per account amplify risks. 2/ Steps to Reconstruct PnL: We approached this as a data pipeline challenge, avoiding risky codebase changes. - Data Collection: Fetch trades, fees, and Binance Kline data. - Mapping: Group data by tokens for efficient processing. - PnL Calculations: 1. Realized PnL: Closed trades minus fees. 2. Unrealized PnL: Position size × (Mark Price - Entry Price). - Aggregation: Sum PnL across user accounts. 3/ Architecture highlights: - Time-series reconstruction: Restore accurate trade timelines. - Parallel processing: Distribute workloads across trading pairs. - Cumulative validation: Track PnL over trade sets for accuracy. 4/ Implementation: - Simultaneously process data for each trading pair, reducing overhead. - Use historical snapshots to rebuild missing PnL states without inconsistencies. - Validate trade sets by tracking cumulative PnL for accuracy and completeness. 5/ Real results: - Accurate reconstruction of historical PnL records. - Processing time reduced from 5 hours to minutes per account. - Designed to handle multiple trading pairs and large datasets seamlessly. Check out the full read: https://lnkd.in/gFMRM8uS #dataengineering #CryptoFinance #dwarves #Software — Dwarves Notes (https://memo.d.foundation/) combine our team’s collective know-hows, R&Ds, and operation approaches. Connect and learn alongside other tech fellows: - Discord: discord.gg/dwarvesv - Github: https://lnkd.in/gZZ2eZMu - Website: https://d.foundation

2

Thank you for sharing this, it’s incredibly timely and valuable! 🙌 Chip Huyen’s work has consistently been a guiding light for practitioners in the ML and AI space, and her new book, AI Engineering: Building Applications with Foundation Models, is exactly what’s needed to navigate the increasingly complex landscape of production-grade Generative AI systems. 📚✨ The trade-offs between open vs. closed source, large vs. small models, and cost vs. performance are central to every engineering decision today. This book seems to tackle these challenges head-on, offering actionable insights for building real-world systems that balance scalability, reliability, and efficiency. The focus on foundation models and their applications is especially exciting, as these models are reshaping how we think about modularity, retraining, and inference pipelines. 🚀 Her previous book, Designing Machine Learning Systems, is already a cornerstone for engineers striving to deploy ML at scale, and this new release is poised to do the same for Generative AI. From choosing the right infrastructure to leveraging frameworks for model serving and monitoring, having such a comprehensive guide will help demystify many of the technical and operational hurdles. The addition of the bi-weekly book reading sessions is a brilliant idea! These sessions will provide a much-needed platform to dive deep into the nuances of AI Engineering—from understanding the trade-offs of foundation model architectures to learning best practices for deployment in production. It's also a great way to foster community learning and exchange ideas on tackling real-world challenges. 🧠💡 Looking forward to seeing how this initiative shapes up! Thanks again for sharing and for creating such meaningful opportunities to learn and grow together. Let’s build production-ready AI systems that are not only cutting-edge but also robust and scalable. 🙌 #AIEngineering #LLMs #MachineLearning #MLOps #GenerativeAI #ProductionAI

1

𝐃𝐚𝐲 7 𝐈𝐧 𝐆𝐞𝐧𝐀𝐈: 𝐄𝐱𝐩𝐥𝐨𝐫𝐢𝐧𝐠 𝐁𝐄𝐑𝐓'𝐬 𝐂𝐨𝐫𝐞 𝐚𝐧𝐝 𝐅𝐢𝐧𝐞-𝐓𝐮𝐧𝐢𝐧𝐠 Day 7 of my Generative AI journey was packed with insights as I wrapped up the BERT paper, "𝐁𝐄𝐑𝐓: 𝐏𝐫𝐞-𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐨𝐟 𝐃𝐞𝐞𝐩 𝐁𝐢𝐝𝐢𝐫𝐞𝐜𝐭𝐢𝐨𝐧𝐚𝐥 𝐓𝐫𝐚𝐧𝐬𝐟𝐨𝐫𝐦𝐞𝐫𝐬 𝐟𝐨𝐫 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐔𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠" by Devlin et al. I also delved into Umair Jamil’s video, "𝐁𝐄𝐑𝐓 𝐞𝐱𝐩𝐥𝐚𝐢𝐧𝐞𝐝: 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠, 𝐈𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞, 𝐁𝐄𝐑𝐓 𝐯𝐬 𝐆𝐏𝐓/𝐋𝐋𝐌, 𝐅𝐢𝐧𝐞-𝐭𝐮𝐧𝐢𝐧𝐠, [𝐂𝐋𝐒] 𝐭𝐨𝐤𝐞𝐧", which added another layer of clarity to my understanding. ( Check it out: https://lnkd.in/dGZg6mQw ) 𝐔𝐧𝐩𝐚𝐜𝐤𝐢𝐧𝐠 𝐭𝐡𝐞 𝐁𝐄𝐑𝐓 𝐏𝐚𝐩𝐞𝐫 The architecture and pre-training strategies in BERT are a masterclass in optimization and versatility. Here’s what stood out: → [𝐂𝐋𝐒] 𝐓𝐨𝐤𝐞𝐧: Used for sentence-level tasks, the token aggregates information from the input sequence and is pivotal for classification models. → 𝐌𝐚𝐬𝐤𝐞𝐝 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐢𝐧𝐠 (𝐌𝐋𝐌): This clever approach of masking random tokens teaches BERT to predict context-sensitive words, enabling a bidirectional understanding of language. → 𝐍𝐞𝐱𝐭 𝐒𝐞𝐧𝐭𝐞𝐧𝐜𝐞 𝐏𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧 (𝐍𝐒𝐏): A task that helps BERT understand sentence relationships, making it effective for tasks like Q&A and summarization. 𝐈𝐧𝐬𝐢𝐠𝐡𝐭𝐬 𝐟𝐫𝐨𝐦 𝐔𝐦𝐚𝐢𝐫 𝐉𝐚𝐦𝐢𝐥’𝐬 𝐕𝐢𝐝𝐞𝐨 Umair Jamil’s explanation bridged the gap between theory and practice. He not only demystified BERT’s architecture but also touched on: → 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐓𝐞𝐜𝐡𝐧𝐢𝐪𝐮𝐞𝐬: How large-scale datasets and compute are utilized for pre-training. → 𝐅𝐢𝐧𝐞-𝐓𝐮𝐧𝐢𝐧𝐠: The ease of adapting BERT for tasks like sentiment analysis and named entity recognition by tweaking the final layers. → 𝐂𝐨𝐦𝐩𝐚𝐫𝐢𝐬𝐨𝐧𝐬: Distinguishing BERT from GPT and other LLMs, emphasizing its pre-training methodology and bidirectional nature. And that was it for today 🤷‍♂️ 𝐖𝐡𝐚𝐭’𝐬 𝐍𝐞𝐱𝐭? Now it’s time to put theory into practice. I'll try implementing the BERT tomorrow. #GenAI #BERT_learnings

10

3 Comments

In context of debate if large context #LLM models will eventually negate the need for #RAG, an article from Databricks has some good insights - • Longer contexts boost RAG accuracy... to a point • Performance often peaks at 32k-64k tokens, then declines • Top models (GPT-4, Claude 3.5) maintain consistency at scale • Others show unique failure modes (e.g., copyright concerns, summarization instead of Q&A) • Optimal context size varies by model and task • Lack of long-context training data may explain some issues Key takeaway: Long context and RAG are synergistic, but LLMs need refinement to fully leverage extended contexts. https://lnkd.in/eehdKtCm

17

Discover advancements in the joint S&P 500/ VIX smile calibration! 💡 Julien Guyon, École nationale des ponts et chaussées joined us at QuantMinds International to share his research on joint S&P 500 / VIX smile calibration, its importance in pricing and hedging derivatives, and the challenges of fitting both options using parametric models. He also dived into the profound impact of machine learning and AI on quant finance! Does its importance match that of the industrial revolution? Find out more 👉 http://spr.ly/6048tdk1e #QuantMinds #quantfinance #ML #AI #hedging

80

1 Comment

That says it all: “Sometimes, the most impactful research isn't about creating shiny new models. It's about saying the obvious thing out loud…. even if your work doesn't feel "novel" enough, your real-world experiences matter. Don't be afraid to state the obvious - it might just be what everyone's thinking, but no one's saying! ” 👉 It is very true, and proven in successful AI practices, that we can built excellent value using AI techniques on the same use case where others claim data quality issues, lack of resources, need for elaborate DL architectures, etc etc. 👉 In the end, experience, understanding of business processes and data, creative thinking, and vision are the key ingredients for success. #AI New York University

33

3 Comments

𝐁𝐫𝐢𝐥𝐥𝐢𝐚𝐧𝐭 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐢𝐧𝐭𝐞𝐫𝐚𝐜𝐭𝐢𝐯𝐞 𝐚𝐧𝐚𝐥𝐲𝐬𝐢𝐬 𝐚𝐧𝐝 𝐌𝐋 𝐚𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬 The architecture is simply brilliant - e2b runtime - LLM middle layer- - notebook style rendering Massive applications. - fully scalable. - open source - customize - build - integrate Next up - prompt-driven iterative analytics and machine learning. Now exploring ways to incorporate similar solution into REX (rex.tigzig.com / realtime.tigzig.com - my open source analytics toolkit), to provide interactive analysis and machine learning. Check out the video demo on X (I actually saw it on X first): https://lnkd.in/gQXqMRiZ App & Code base : https://lnkd.in/eGmyXh-q

2

🚀[#LLMs vs #XgBoost]🚀 These are interesting reads and open up avenues of how the future of model.fit() looks like for tabular data. Where this helps ? 📍Feature engineering: LLMs understand the context and why not they are trained against huge corpus of text, videos and images. What’s needed is to take tabular data and bring them to a serialised format for LLMs to make sense of it. No more handcoding of features and wonder what works best for downstream task 📍Model.fit(): LLMs can then be used for embedding generations and let neural nets do the magic for you 📍Interpretability: The very same LLMs can explain why they make a particular decision. After all they are digital assistants 📍Quick updates: Unlike xgboost, LLMs can be used for faster model update Link: https://lnkd.in/dhZV-3C3

3

Bitcoin: A Unique Diversifier - Is bitcoin a 'risk on' or 'risk off' asset ? Read this insightful paper from the team at Blackrock to find out their views ! #bitcoin #digitalassets #blackrock

6