requirements.txt

# =============================================================================
# Verified with (Python 3.11), last full run 2026-06-07 — the course executes
# clean on this exact set (72 pass / 3 intentional Debug-me xfail / 0 fail):
#   numpy 2.4.6 · pandas 3.0.3 · matplotlib 3.10.9 · scikit-learn 1.8.0
#   statsmodels 0.14.6 · scipy 1.17.1 · requests 2.34.2 · jupyterlab 4.5.7
# Floors stay open ("modern tools, modern habits" — see README), but each line
# is now capped below the NEXT major release. That keeps you on the latest
# patch/minor while blocking a future breaking major from silently breaking a
# notebook (e.g. the pandas copy-on-write change that needed a fix in NB 37).
# For a fully frozen baseline instead, run: pip freeze > requirements.lock
# =============================================================================

# Core libraries used across the course
numpy>=1.24,<3
pandas>=2.0,<4
matplotlib>=3.9,<4     # 3.9+ required: notebooks use ax.boxplot(tick_labels=...)
seaborn>=0.13,<1       # statistical charts in one line (NB 9, fast-track NB 7)
scikit-learn>=1.3,<2
scipy>=1.10,<2        # statistics, safety stock, rank ensembles (NB 10, 35–38)
jupyterlab>=4.0,<5
ipywidgets>=8.0,<9    # interactive sliders in the model->money cells (NB 35, 36, 38)

# Notebook 12 — real HTTP calls
requests>=2.30,<3

# Notebook 11 — time series & forecasting
statsmodels>=0.14,<1

# Notebook 22 — testing inside the project
pytest>=8.0,<10

# Notebook 23 — optional in-process scheduling
# schedule>=1.2

# --- Optional LLM providers (see notebook A1 — the providers guide) ---
# Every AI notebook works offline with the built-in MockLLM. Uncomment the
# provider(s) you actually plan to use.
# openai>=1.0                       # OpenAI: gpt-4o-mini, gpt-4o, text-embedding-3-small
# anthropic>=0.30                   # Anthropic: Claude haiku / sonnet / opus
# google-generativeai>=0.7          # Google: Gemini Flash / Pro
# ollama>=0.3                       # Local: any model you `ollama pull` (needs Ollama server)
# sentence-transformers>=2.7        # Local embeddings (NB 18 — embeddings & retrieval)

# --- Optional vector stores & RAG/agent frameworks (Module 5 appendices A2-A3) ---
# faiss-cpu>=1.7
# chromadb>=0.5
# qdrant-client>=1.9
# langchain>=0.2
# llama-index>=0.10
# haystack-ai>=2.0

# --- Optional deep learning (Module 4 appendices A1-A3) ---
# torch>=2.2                        # install: pip install torch --index-url https://download.pytorch.org/whl/cpu
# transformers>=4.40
# peft>=0.10                        # LoRA / parameter-efficient fine-tuning

# --- Optional tabular foundation model (Module 4 appendix A4) ---
# tabpfn>=0.1                       # local TabPFN (PriorLabs)
# tabpfn-client>=0.1                # hosted PriorLabs API

# --- Optional text analytics (Module 12 — topic modeling & sentiment) ---
# Each notebook runs offline via a scikit-learn stand-in; install to see the real library.
# bertopic>=0.16                    # NB 43 — embedding topic models (pulls sentence-transformers, umap-learn, hdbscan)
# stream-topic>=0.1                 # NB 44 — STREAM (AnFreTh/STREAM): unified topic models + evaluation
# vaderSentiment>=3.3               # NB 45 — lexicon/rule-based sentiment
# transformers>=4.40                # NB 45 — transformer sentiment pipelines (also Module 4 A3)

# --- Optional deep tabular learning (Module 13 — DeepTab) ---
# deeptab>=0.1                      # NB 46 — OpenTabular/DeepTab; deep tabular models, sklearn API (pulls torch + lightning)

# --- Optional forecasting libraries (Module 2 appendices A2-A4) ---
# prophet>=1.1
# neuralprophet>=0.8
# sktime>=0.30
# darts>=0.30                       # was published as u8darts; current package name is `darts`
# timesfm>=1.0                      # Google's foundation forecaster
# chronos-forecasting>=1.2          # Amazon's foundation forecaster
# tabpfn-time-series>=0.1           # PriorLabs forecasting model