[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-76144":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":8,"htmlUrl":8,"language":9,"languages":8,"totalLinesOfCode":8,"stars":10,"forks":11,"watchers":12,"openIssues":13,"contributorsCount":13,"subscribersCount":13,"size":13,"stars1d":12,"stars7d":14,"stars30d":15,"stars90d":13,"forks30d":13,"starsTrendScore":16,"compositeScore":17,"rankGlobal":8,"rankLanguage":8,"license":18,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":19,"hasPages":19,"topics":21,"createdAt":8,"pushedAt":8,"updatedAt":22,"readmeContent":23,"aiSummary":24,"trendingCount":13,"starSnapshotCount":13,"syncStatus":25,"lastSyncTime":26,"discoverSource":27},76144,"Deep_Learning_for_Solving_And_Estimating_Dynamic_Economic_Models","sischei\u002FDeep_Learning_for_Solving_And_Estimating_Dynamic_Economic_Models","sischei",null,"Jupyter Notebook",170,39,1,0,6,31,3,4.81,"Other",false,"main",[],"2026-06-12 02:03:40","# Deep Learning for Solving and Estimating Dynamic Models in Economics and Finance\n\n*Deep learning for the recursive, stochastic, high-dimensional dynamic\nmodels that economists actually solve, with all materials open source,\nrunnable, and self-contained.*\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"lecture_script\u002FDeep_Learning_for_Solving_And_Estimating_Dynamic_Economic_Models.pdf\">\u003Cimg src=\"assets\u002Fhero\u002Fdeep_learning_dynamic_models_hero.png\" width=\"95%\" alt=\"Deep Learning for Solving and Estimating Dynamic Models in Economics and Finance\">\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"COURSE_MAP.md\">\u003Cimg alt=\"18 lectures\" src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Flectures-18-1f4e79?style=for-the-badge\">\u003C\u002Fa>\n  \u003Ca href=\"lectures\u002F\">\u003Cimg alt=\"Runnable notebooks\" src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fnotebooks-runnable-1f4e79?style=for-the-badge\">\u003C\u002Fa>\n  \u003Ca href=\"lecture_script\u002FDeep_Learning_for_Solving_And_Estimating_Dynamic_Economic_Models.pdf\">\u003Cimg alt=\"Companion script\" src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fcompanion%20script-PDF-1f4e79?style=for-the-badge\">\u003C\u002Fa>\n  \u003Cbr\u002F>\n  \u003Ca href=\"lectures\u002Flecture_06_agentic_programming\u002F\">\u003Cimg alt=\"AI-coding workshop\" src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAI%20coding-workshop-1f4e79?style=for-the-badge\">\u003C\u002Fa>\n  \u003Ca href=\"#syllabus\">\u003Cimg alt=\"Exercises with solutions\" src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fexercises-with%20solutions-1f4e79?style=for-the-badge\">\u003C\u002Fa>\n  \u003Ca href=\"LICENSE\">\u003Cimg alt=\"Open source, self-paced, no enrollment\" src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fopen%20source-self--paced-1f4e79?style=for-the-badge\">\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Ch3 align=\"center\">\n  Course author: \u003Ca href=\"https:\u002F\u002Fsischei.github.io\u002F\">Simon Scheidegger\u003C\u002Fa>\n\u003C\u002Fh3>\n\u003Cp align=\"center\">\u003Cem>University of Lausanne  ·  Grantham Research Institute, London School of Economics\u003C\u002Fem>\u003C\u002Fp>\n\n\u003Cbr\u002F>\n\n## About this course\n\n**Classical grid methods hit a wall.** Modern macroeconomics, finance,\nand climate economics have outgrown the grid-based numerical methods\nthat dominated a generation ago. Once you add heterogeneous agents,\noverlapping generations, occasionally binding constraints,\ncontinuous-time dynamics, or coupled climate-economic interactions,\nthe state space becomes too large for tensor-product grids and\nclassical methods (projection, value-function iteration, perturbation)\nbreak down. If you are trying to solve models with ten or more state\ndimensions, estimate them, or design policy under parameter uncertainty,\nyou need a different toolbox.\n\n**This course teaches that toolbox.** A coherent set of deep-learning\nmethods built *for* the recursive, stochastic, often high-dimensional\nmodels economists actually solve. The methods work by letting economic\nstructure drive the learning problem: equilibrium conditions, Bellman\nequations, and PDEs become the residual loss, e.g., in Deep\nEquilibrium Nets or Physics-Informed Neural Networks (an unsupervised\nsetup), or they shape the simulator that generates the (input, output)\npairs a deep surrogate or Gaussian process then learns in the standard\nsupervised way. You will build each method from scratch on benchmarks\nwhere the answer is known (Brock–Mirman, cake-eating, Black–Scholes)\nbefore applying them to models where it is not (IRBC, OLG with 56\ncohorts, Krusell–Smith with a continuum, continuous-time heterogeneous\nagents, climate-economic coupling). The course is hands-on by design:\nevery method is paired with runnable Jupyter notebooks that put the\nprinciples in plain sight, so you see exactly how each loss is\nassembled, each gradient is taken, and each equilibrium is solved\nrather than reading about it. By the end you will be able to\nsolve models that were out of reach with classical tools, estimate\nthem when re-solving is too expensive, and design policies that take\nparameter uncertainty seriously.\n\n**Everything is self-contained and open source.** A textbook-length\n[companion script](lecture_script\u002FDeep_Learning_for_Solving_And_Estimating_Dynamic_Economic_Models.pdf),\n**18 paired lectures** with slides and runnable Jupyter notebooks,\nexercises with full solutions, a hands-on workshop on AI coding agents\nas research partners (Lecture 06), and a curated bibliography linking\nout to the underlying papers. There is no enrollment, no cohort, no\ndeadline, just pick the method you need and dig in.\n\n> **A working snapshot, not a definitive survey.** The selection of\n> methods, papers, and implementation choices reflects what I\n> currently find to be some of the most useful entry points for\n> economists and finance researchers starting to work with deep\n> learning in dynamic models, and the bibliography is necessarily\n> incomplete. Treat the material as a practical entry point to study,\n> run, adapt, and question.\n\n> ### 🚀 Start here\n>\n> - **Jump in:** [Lecture 02 — Intro to deep learning](lectures\u002Flecture_02_intro_deep_learning\u002FREADME.md)\n> - **New to Python?** Begin with the [Python primer (Lecture 01)](lectures\u002Flecture_01_python_primer\u002FREADME.md)\n> - **Want the panoramic view?** Open the [course map](COURSE_MAP.md)\n\n## What you will learn\n\nThis course teaches a coherent set of deep-learning methods for the\nrecursive, stochastic, often high-dimensional models that show up in\nmodern macroeconomics, asset pricing, and climate-economic policy work.\nFive capabilities, each motivated below.\n\n### 1. Solving recursive equilibrium models with neural networks\n\nMost quantitative macro models reduce to functional equations (Euler\nequations, Bellman equations, market-clearing conditions) that\nclassical methods (projection, value-function iteration, perturbation)\nstruggle with once the state space gets large or the policy is\nnonsmooth. **Deep Equilibrium Nets (DEQNs)** parameterize the policy\nor value function with a neural network and minimize the\nequilibrium-equation residuals directly via stochastic gradient descent,\nsidestepping a curse-of-dimensionality grid. The companion **Physics-Informed\nNeural Networks (PINNs)** do the same for continuous-time models: the\nloss is the residual of a Hamilton–Jacobi–Bellman equation, automatic\ndifferentiation supplies the derivatives, and there is no mesh.\nYou will build both end-to-end on benchmarks where the answer is known\n(Brock–Mirman, cake-eating, Black–Scholes) and then on models where it\nis not (IRBC, OLG with 56 cohorts, Krusell–Smith with a continuum of\nagents, continuous-time heterogeneous agents).\n\n### 2. Surrogates, Gaussian processes, and Bayesian active learning\n\nMany calibration, estimation, and policy-evaluation tasks call the\nunderlying model thousands or millions of times. A **deep surrogate\nmodel** replaces that expensive call with a cheap, differentiable\nneural network trained on a few hundred or thousand simulator outputs.\nA **Gaussian process (GP)** does the same with built-in uncertainty\nquantification, which lets **Bayesian active learning (BAL)** pick the\nnext training point optimally instead of throwing samples at a\nhypercube. We then push GPs to high dimension via **active subspaces**\nand **deep kernel learning**, and use them inside value-function\niteration (ASGP-VFI) as a competitor to DEQNs.\n\n### 3. Structural estimation via simulated method of moments\n\nOnce a deep surrogate is in place, **simulated method of moments\n(SMM)** estimation becomes a small optimization over the surrogate\nrather than a brutal repeated re-solve of the structural model. You\nwill run single- and joint-parameter SMM on a deep surrogate of\nBrock–Mirman and see how the estimator behaves under realistic noise\nand identification challenges.\n\n### 4. Deep UQ and Pareto-improving climate policy\n\nIntegrated assessment models (DICE, CDICE) carry parameters whose\ntrue values are deeply uncertain, equilibrium climate sensitivity\nbeing the textbook example. Plugging point estimates in and reading\noff a single social cost of carbon is misleading; averaging the\nuncertainty out before optimization is worse, because the policy you\nwould choose under expected damages is generally not the policy you\nwould choose if you took the tail risk seriously.\n\nThe course teaches a complete pipeline that addresses this directly.\nWe solve a stochastic IAM with DEQNs under Epstein–Zin preferences,\nbuild GP surrogates for the quantities of interest with Bayesian\nactive learning, and run global sensitivity analysis (Sobol, Shapley\neffects) to localize where the policy is actually sensitive to which\nparameters. On top of that surrogate we then **design constrained\nPareto-improving carbon-tax policies**: tax paths that, for every\nplausible parameter draw (or every cohort, or every generation), leave\nno agent worse off than the business-as-usual baseline while strictly\nimproving welfare for at least one. This turns \"what should the\ncarbon tax be?\" from a single number computed under a single\ncalibration into a defensible policy menu that respects who bears the\nrisk and who benefits, *without* averaging the uncertainty away.\n\n### 5. An AI-assisted research-coding workflow\n\nModern empirical and computational economics benefits enormously from\nusing AI coding agents (Claude Code) as research partners, but only\nwhen the workflow is set up deliberately. **Lecture 06** is a\nhands-on workshop that teaches the orientation, prompt patterns,\nproject memory (`CLAUDE.md`), custom skills, subagents, and hooks\nthat turn an LLM from a clever autocomplete into a real research\ncollaborator, paired with twelve self-paced exercises so you walk\nout with reusable templates rather than slideware.\n\n## How to use this course\n\nDifferent readers come in with different goals, so pick the entry point\nthat fits yours:\n\n- 🚀 **I want a guided start.** Open the\n  [Python primer (Lecture 01)](lectures\u002Flecture_01_python_primer\u002FREADME.md)\n  if you need it, then follow the **Complete path** in\n  [`COURSE_MAP.md`](COURSE_MAP.md). It walks through all 18 lectures\n  in their natural order.\n- 🎯 **I have a specific topic in mind.** Jump straight to the\n  **syllabus** below.\n- 🧪 **I want the research-workflow training first.** Jump to\n  [Lecture 06, agentic programming](lectures\u002Flecture_06_agentic_programming\u002FREADME.md),\n  then come back to the rest of the sequence.\n- 📖 **I want a textbook.** Read the chapter-based\n  [companion script](lecture_script\u002FDeep_Learning_for_Solving_And_Estimating_Dynamic_Economic_Models.pdf); each chapter\n  links to one or more lectures via\n  [`script_to_lectures.md`](lecture_script\u002Fscript_to_lectures.md).\n\nFor each lecture, the workflow is the same:\n\n1. read the relevant chapter or section of the script;\n2. step through the lecture's slide deck (under `slides\u002F`);\n3. run the lecture's notebooks under `code\u002F` (numbered in suggested\n   order; files ending in `_Exercises_Blanks.ipynb` \u002F\n   `_Exercises_Solutions.ipynb` are paired exercise\u002Fsolution sets).\n\nEvery long-running notebook exposes a `RUN_MODE` switch near the top with\nthree values: `\"smoke\"` (CPU-bounded, runs in minutes for a sanity check\nor CI), `\"teaching\"` (laptop figures, intermediate fidelity), and\n`\"production\"` (full reproduction, published-figure quality). Each\nnotebook also fixes a `SEED` for reproducibility.\n\n## Syllabus\n\n| If you want to learn… | Read | Notebooks |\n|---|---|---|\n| **Python warm-up** (skip if you write Python every day) | [Lecture 01](lectures\u002Flecture_01_python_primer\u002FREADME.md) | Jupyter, basic data structures, NumPy, plotting, classes |\n| **Deep-learning fundamentals** (training, generalization, sequence models) | [Lecture 02](lectures\u002Flecture_02_intro_deep_learning\u002FREADME.md) | MLP, LSTM, Transformer on Edgeworth cycles, double descent, Genz approximations |\n| **Deep Equilibrium Nets (DEQNs)**, the central method | [Lecture 03](lectures\u002Flecture_03_deep_equilibrium_nets\u002FREADME.md) | Brock–Mirman (deterministic, stochastic), Fischer–Burmeister constraints, six loss kernels |\n| **Large-scale nonlinear DSGE** (IRBC) | [Lecture 04](lectures\u002Flecture_04_irbc_with_deqns\u002FREADME.md) | International real business cycle with DEQNs |\n| **Architecture search and loss balancing** (NAS, ReLoBRaLo) | [Lecture 05](lectures\u002Flecture_05_nas_loss_normalization\u002FREADME.md) | Random search, Hyperband, ReLoBRaLo, SoftAdapt, GradNorm |\n| **Agentic programming** (AI coding agents as research partners) | [Lecture 06](lectures\u002Flecture_06_agentic_programming\u002FREADME.md) | Claude Code workflow, prompts, project memory, custom skills, subagents, hooks, plus a 12-exercise workshop |\n| **Automatic differentiation for DEQNs** | [Lecture 07](lectures\u002Flecture_07_autodiff_for_deqns\u002FREADME.md) | Lagrangian primitives, two-tape gradients, IRBC autodiff |\n| **OLG with DEQNs** | [Lecture 08](lectures\u002Flecture_08_olg_models_deqns\u002FREADME.md) | Analytic OLG, 56-cohort benchmark, Fischer–Burmeister borrowing constraints |\n| **Heterogeneous agents and Young's method** | [Lecture 09](lectures\u002Flecture_09_heterogeneous_agents_youngs_method\u002FREADME.md) | Young's histogram, Krusell–Smith, continuum-of-agents DEQN |\n| **Sequence-space DEQNs** | [Lecture 10](lectures\u002Flecture_10_sequence_space_deqns\u002FREADME.md) | Brock–Mirman, IRBC, Krusell–Smith with shock-history inputs |\n| **Physics-informed neural networks (PINNs)** | [Lecture 11](lectures\u002Flecture_11_pinns\u002FREADME.md) | ODE \u002F PDE PINNs, soft vs hard BCs, cake-eating HJB, Black–Scholes |\n| ↳ **Continuous-time HA, theory** | [Lecture 12](lectures\u002Flecture_12_continuous_time_ha_theory\u002FREADME.md) | HJB, Kolmogorov-forward, master equation, Ito calculus |\n| ↳ **Continuous-time HA, numerics** | [Lecture 13](lectures\u002Flecture_13_continuous_time_ha_numerics\u002FREADME.md) | Achdou–Han–Lasry–Lions–Moll finite-difference scheme, PINN for HJB-KFE, continuous-time Aiyagari |\n| **Surrogates, Gaussian processes, deep kernels** | [Lecture 14](lectures\u002Flecture_14_surrogates_and_gps\u002FREADME.md) | Surrogate primer, GP regression, BAL, active subspaces, deep kernel learning, GP-VFI |\n| **Structural estimation via SMM** | [Lecture 15](lectures\u002Flecture_15_structural_estimation_smm\u002FREADME.md) | Brock–Mirman SMM (single- and joint-parameter) on a deep surrogate |\n| **Climate economics and IAMs (DICE, CDICE)** | [Lecture 16](lectures\u002Flecture_16_climate_economics_iams\u002FREADME.md) | DICE \u002F CDICE simulation, deterministic and stochastic CDICE-DEQN |\n| ↳ **Deep UQ and Pareto-improving carbon-tax design** | [Lecture 17](lectures\u002Flecture_17_deep_uq_pareto_improving_policy\u002FREADME.md) | GP surrogates, Bayesian active learning, Sobol \u002F Shapley, constrained Pareto-improving carbon-tax rules |\n| **Synthesis, when to use which method** | [Lecture 18](lectures\u002Flecture_18_course_wrap_up\u002FREADME.md) | Decision guide and outlook |\n\nFor the full table including compute and time budgets, prerequisites,\nand the visual prerequisite diagram, see\n[`COURSE_MAP.md`](COURSE_MAP.md).\n\n## Setup\n\nNotebooks run on **Python 3.10+**. Two reproducible setups:\n\n```bash\n# pip\npip install -r requirements.txt\n\n# conda\nconda env create -f environment.yml\nconda activate dlef\n```\n\nMain dependencies: NumPy, SciPy, pandas, Matplotlib, scikit-learn,\nTensorFlow ≥ 2.15, PyTorch ≥ 2.0, JAX (selected notebooks), GPyTorch\nand BoTorch (Lecture 13).\n\n## Repository at a glance\n\n```\n.\n├── README.md             ← you are here\n├── COURSE_MAP.md         ← detailed map, learning paths, prerequisite diagram\n├── lectures\u002F             ← 18 lecture folders (lecture_XX_*)\n│   └── lecture_*\u002F\n│       ├── README.md         summary, slides, code, prerequisites, readings\n│       ├── slides\u002F           PDFs and .tex sources\n│       ├── code\u002F             notebooks, supporting .py modules, data files\n│       └── figures\u002F          (optional) lecture-specific figure assets\n├── lecture_script\u002F       ← textbook-length companion script\n├── readings\u002F             ← per-lecture link guides + bibliography.bib\n└── assets\u002F               ← hero figure, generated figures, attributions\n```\n\n## Glossary\n\nThe script's Appendix A is the canonical glossary. A grep-able copy\nlives at [`lecture_script\u002Fglossary.md`](lecture_script\u002Fglossary.md).\n\n## Readings and copyright\n\nMost readings are journal articles, working papers, or copyrighted\nbooks. The public repository **links** to publishers, DOIs, arXiv, or\nauthor pages rather than redistributing PDFs. Per-lecture link guides\nlive under\n[`readings\u002Flinks_by_lecture\u002F`](readings\u002Flinks_by_lecture\u002F);\nthe full bibliography is in\n[`readings\u002Fbibliography.bib`](readings\u002Fbibliography.bib).\n\nCourse author: **[Simon Scheidegger](https:\u002F\u002Fsischei.github.io\u002F)** (University of Lausanne).\nCode is MIT-licensed; text, slides, script, and figures are CC0\n(see [`LICENSE`](LICENSE) for both).\n\n## Citation\n\nIf this work was useful in your research, please cite the arXiv\nmanuscript (preferred) or the SSRN version:\n\n**arXiv (preferred):**\n\n```bibtex\n@misc{scheidegger2026deeplearningsolvingestimating,\n  title         = {Deep Learning for Solving and Estimating Dynamic Models in Economics and Finance},\n  author        = {Simon Scheidegger},\n  year          = {2026},\n  eprint        = {2605.14493},\n  archivePrefix = {arXiv},\n  primaryClass  = {econ.GN},\n  url           = {https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.14493}\n}\n```\n\n**SSRN:**\n\n```bibtex\n@article{scheidegger_2026_ssrn,\n  title   = {Deep Learning for Solving and Estimating Dynamic Models in Economics and Finance},\n  author  = {Scheidegger, Simon},\n  year    = {2026},\n  month   = {5},\n  doi     = {10.2139\u002Fssrn.6758340},\n  url     = {https:\u002F\u002Fssrn.com\u002Fabstract=6758340},\n  journal = {Available at SSRN 6758340},\n  note    = {Posted 13 May 2026}\n}\n```\n\n## Errata, contributions, and contact\n\nQuestions, corrections, and pull requests are welcome on\n[GitHub](https:\u002F\u002Fgithub.com\u002Fsischei\u002FDeep_Learning_for_Solving_And_Estimating_Dynamic_Economic_Models).\nBy contributing you agree that your contribution is licensed under the\nsame terms as this repository.\n","该项目专注于利用深度学习解决和估计经济学与金融学中的递归、随机及高维动态模型。其核心功能包括通过经济结构驱动学习过程，例如将均衡条件、贝尔曼方程或偏微分方程转化为残差损失，适用于Deep Equilibrium Nets或物理信息神经网络等无监督设置；或者这些方法也可以指导生成器产生（输入，输出）对，供深度代理或高斯过程在标准监督学习框架下学习。此项目特别适合那些需要处理超过十个状态维度的复杂经济模型的研究者使用，如涉及异质性主体、世代交叠、偶尔约束绑定、连续时间动态或耦合气候-经济互动等问题时。所有材料均为开源，并且提供了可运行的Jupyter笔记本、18个讲座视频以及配套讲义，非常适合自学。",2,"2026-06-11 03:54:39","CREATED_QUERY"]