دانلود کتاب یادگیری ماشین برای سری‌های زمانی با پایتون

In March 2020, Target’s machine learning models, built by worldclass data science teams, told the company to keep stocking jeans and party dresses. Customers were busy clearing the shelves of toilet paper, sweatpants, and cleaning supplies. Apparel comp sales fell more than 30 percent in the worst weeks of March; electronics and household essentials surged 20 to 45 percent in the other direction. At Amazon, the algorithms that had balanced millions of SKUs were overwhelmed: household essentials vanished from the site, and the company suspended inbound shipments of anything it classified as non-essential just to free up warehouse capacity. At Walmart, stockouts climbed across 4,700 U.S. stores faster than the planning systems could adjust, and U.S. inventory finished the quarter down nearly five percent year-over-year. Within weeks, billions of dollars of forecasting infrastructure went from a competitive advantage to a source of systematic error. Twenty years earlier, Nike had taken a $100 million hit to quarterly sales, and a 20 percent stock drop the day it was announced, after its newly installed i2 demand planner over-ordered Air Garne_ sneakers and under-ordered Air Jordans. In December 2011, Best Buy canceled online holiday orders days before Christmas, weeks after it had accepted payment, blaming demand that had outrun stock. In August 2020, California’s grid operators ordered rolling blackouts because their solar-generation forecasts had not learned the negative correlation between sunset and air-conditioning load during a heatwave. Different industries, different decades, the same pa_ern: smart teams with state-of-the-art models, blindsided by the way time series actually behaves in production. This book is about why that keeps happening and how to build systems that hold up when it does. The short answer is that time series breaks the assumption almost every other machine learning method depends on, that observations are independent. Today’s value is a function of yesterday’s. Shuffle the rows and the signal disappears. Validate with a random split and you train on the future, test on the past, and ship a model that looks great in the notebook and fails the first time it meets a pattern it has not memorized. Add the fact that most production forecasting pipelines are 5 percent algorithm and 95 percent data engineering, calibration, monitoring, and drift handling (what Chapter 1 calls the Iceberg Problem), and you have a discipline where the gap between a working prototype and a working system is wider than in almost any other corner of machine learning. Across twelve chapters, we work through the full stack a modern forecasting practitioner needs: diagnostics, classical baselines, machine learning, deep learning, hierarchical, and multivariate models, conformal prediction intervals, time-series foundation models, production deployment, anomaly detection, and classification, and the awkward cases (intermittent demand, censored events) that most introductory books skip. The thread holding it together is the production gap. Every chapter asks two questions: does the model fit? and will the model survive a Tuesday?