| --- |
| pretty_name: "Plan B — Contrarian Deep-Value US Equities & AI 10-K Analysis" |
| license: cc-by-nc-4.0 |
| language: |
| - en |
| size_categories: |
| - n<1K |
| task_categories: |
| - tabular-classification |
| - tabular-regression |
| - text-generation |
| - time-series-forecasting |
| tags: |
| - finance |
| - equities |
| - stocks |
| - sec-edgar |
| - xbrl |
| - value-investing |
| - fundamental-analysis |
| - 10-k |
| - form-4 |
| - insider-trading |
| - financial-nlp |
| configs: |
| - config_name: daily_rankings |
| default: true |
| data_files: |
| - split: train |
| path: daily_rankings.jsonl |
| - config_name: ai_10k_memos |
| data_files: |
| - split: train |
| path: ai_10k_memos.jsonl |
| - config_name: insider_trades |
| data_files: |
| - split: train |
| path: insider_trades.jsonl |
| - config_name: scan_stats |
| data_files: |
| - split: train |
| path: scan_stats.jsonl |
| --- |
| |
| # Plan B — Contrarian Deep-Value US Equities & AI 10-K Analysis |
|
|
| **Plan B** is a daily quantitative and fundamental research dataset covering deeply discounted, financially resilient, and heavily beaten-down US operating companies across the **NYSE** and **Nasdaq**. |
|
|
| Every trading day, Plan B screens the entire US equity universe (~6,000 SEC XBRL filers) using primary **SEC EDGAR XBRL frames**, **3-year weekly price series**, **SEC Form 4 open-market insider transactions**, **news sentiment**, and **LLM-synthesized 10-K/20-F deep-value investment memos**. |
|
|
| --- |
|
|
| ## Dataset Snapshot |
|
|
| | Metric | Value | |
| | :--- | :--- | |
| | **Coverage Window** | `2026-11-02` → `2026-11-06` (5 daily scans) | |
| | **Unique Shortlisted Companies** | 54 NYSE/Nasdaq tickers | |
| | **Daily Ranking Observations** | 200 rows (68 features per row) | |
| | **Structured AI 10-K/20-F Memos** | 16 filings analyzed | |
| | **Open-Market Form 4 Insider Trades** | 305 verified transactions | |
| | **Update Frequency** | Daily (automated post-market scan) | |
| | **License** | `CC-BY-NC-4.0` (Research & Non-Commercial) | |
|
|
| --- |
|
|
| ## Dataset Structure & Files |
|
|
| Each table is provided in both **CSV** (ideal for Kaggle, Pandas, R, and DuckDB) and **JSON Lines (`.jsonl`)** (ideal for Hugging Face `datasets`, Spark, and LLM pipelines). |
|
|
| | Table / Config | Files | Primary Key | Description | |
| | :--- | :--- | :--- | :--- | |
| | **`daily_rankings`** *(default)* | `daily_rankings.csv`<br>`daily_rankings.jsonl` | `(date, ticker)` | Daily ranked contrarian shortlist (~40 companies/day) with composite scores, valuation ratios, SEC XBRL balance-sheet & cash-flow fundamentals, insider totals, and AI 10-K verdict summary. | |
| | **`ai_10k_memos`** | `ai_10k_memos.csv`<br>`ai_10k_memos.jsonl` | `(ticker, accession)` | Skeptical value-investor memos generated from Item 1 (Business), Item 1A (Risk Factors), and Item 7 (MD&A) of each company's latest SEC Form 10-K or 20-F. | |
| | **`insider_trades`** | `insider_trades.csv`<br>`insider_trades.jsonl` | `(ticker, owner, kind, date, shares, form4_url)` | Open-market insider purchases (`P`) and sales (`S`) parsed directly from SEC Form 4 XML filings over the trailing 180 days (excluding option grants/awards). | |
| | **`scan_stats`** | `scan_stats.csv`<br>`scan_stats.jsonl` | `date` | Daily universe screening funnel: total filers, usable XBRL fundamentals, priced equities, and shortlisted count. | |
| |
| --- |
| |
| ## Screening & Ranking Methodology |
| |
| Plan B follows a strict Graham–Klarman–Buffett deep-value discipline. Financial institutions (SIC 6000–6799: banks, insurers, REITs, SPACs) are excluded so balance-sheet ratios remain comparable across operating businesses. |
| |
| ### 1. Hard Quantitative Screen (`Cheap + Safe + Hated`) |
| A company must pass **all three** pillars simultaneously: |
| - **Cheap:** P/E ≤ 12, **or** EV/EBIT ≤ 8, **or** trading below Net Current Asset Value (`NCAV / Market Cap ≥ 1.0`, a classic Benjamin Graham net-net). |
| - **Safe:** Positive stockholders' equity, current ratio ≥ 1.2, total liabilities ≤ 2.0× equity (or net cash position), 3-year share dilution ≤ 25%, and positive operating cash flow if net income is negative. |
| - **Hated:** Share price down at least **40%** from its 3-year weekly high (`dd_3y ≤ -0.40`), with market cap ≥ $50M and share price ≥ $1.00. |
| |
| ### 2. Composite Score (`0–100`) |
| - **Base Weighting:** 40% **Cheap** (`cheap_score`) + 30% **Safe** (`safe_score`) + 30% **Hated** (`hated_score`, blending price drawdown and headline negativity `news_neg`). |
| - **Bonuses:** Net open-market Form 4 insider buying (up to +12 pts) and 3-year share count reduction via buybacks (+4 pts). |
| - **Penalties:** Share dilution, negative free cash flow, accrual divergence (net income significantly exceeding operating cash flow or operating income), one-off profits following consecutive loss years, and high goodwill/M&A intensity. |
| - **AI 10-K Adjustment:** After reading the 10-K filing, the composite score is adjusted by the AI analyst's verdict (`CONTRARIAN BUY CANDIDATE`, `WATCH`, or `AVOID`) and conviction (`0–100`). |
|
|
| --- |
|
|
| ## Quickstart Usage |
|
|
| ### Python (`datasets` — Hugging Face) |
| ```python |
| from datasets import load_dataset |
| |
| # Load default config (daily_rankings) |
| rankings = load_dataset("elsiddik/planb-deep-value", "daily_rankings", split="train") |
| print(rankings.to_pandas()[["date", "rank", "ticker", "score", "pe", "ev_ebit", "dd_3y", "ai_verdict"]].head()) |
| |
| # Load structured 10-K AI memos |
| memos = load_dataset("elsiddik/planb-deep-value", "ai_10k_memos", split="train") |
| |
| # Load Form 4 open-market insider trades |
| insiders = load_dataset("elsiddik/planb-deep-value", "insider_trades", split="train") |
| ``` |
|
|
| ### Python (`pandas` — Kaggle / Local CSV) |
| ```python |
| import pandas as pd |
| |
| df = pd.read_csv("daily_rankings.csv") |
| memos = pd.read_csv("ai_10k_memos.csv") |
| insiders = pd.read_csv("insider_trades.csv") |
| |
| # Filter highest-conviction ideas with net cash and insider buying |
| top_ideas = df[(df["net_cash"] > 0) & (df["insider_buy_value"] > 0)].sort_values(["date", "rank"]) |
| print(top_ideas[["date", "ticker", "name", "score", "pe", "ev_ebit", "net_cash_pct", "ai_verdict"]].head(10)) |
| ``` |
|
|
| --- |
|
|
| ## Data Dictionary |
|
|
| ### 1. `daily_rankings` (68 Columns) |
| |
| | Column | Type | Description | |
| | :--- | :--- | :--- | |
| | `date` | `string` | Scan date (`YYYY-MM-DD`). | |
| | `rank` | `int` | Daily rank (`1` = highest composite score). | |
| | `ticker`, `name`, `cik`, `exchange`, `sector` | `string`/`int` | Ticker symbol, SEC registrant name, SEC CIK identifier, exchange (`NYSE`/`Nasdaq`), and SIC industry description. | |
| | `score` | `float` | Final Plan B composite score (`0–100`). | |
| | `cheap_score`, `safe_score`, `hated_price_score`, `hated_score` | `float` | Normalized sub-scores (`0.0–1.0`) for valuation, balance-sheet safety, price drawdown, and combined hated score. | |
| | `news_neg`, `hate_why` | `float`/`string` | Headline negativity score (`0.0` loved → `1.0` hated) and short explanation of market pessimism. | |
| | `first_seen`, `return_since_first` | `string`/`float` | Date the ticker first entered the shortlist and cumulative price return since `first_seen`. | |
| | `bonus`, `flags` | `list` / `string` | Score bonuses (e.g., insider buying, shrinking shares) and accounting/risk flags. | |
| | `price`, `mcap`, `ev` | `float` | Latest share price (USD), market capitalization (USD), and enterprise value (`Market Cap + Debt - Cash`, USD). | |
| | `dd_3y`, `ret_1y`, `high_3y`, `high_52w`, `low_52w`, `off_low` | `float` | Drawdown from 3-year high (`-0.75` = down 75%), 1-year return, 3-year/52-week highs & lows, and percentage above 52-week low. | |
| | `pe`, `ev_ebit`, `pb` | `float` | Price-to-Earnings, Enterprise Value-to-EBIT (capped at `0` when EV is negative), and Price-to-Book ratio. | |
| | `earnings_yield`, `ebit_ev`, `fcf_yield`, `book_yield` | `float` | Inverse valuation yields: `NI / MCap`, `EBIT / EV`, `FCF / MCap`, and `Equity / MCap`. | |
| | `ncav`, `netnet`, `net_cash`, `net_cash_pct` | `float` | Net Current Asset Value (`CA - Total Liabilities`), `NCAV / MCap`, Net Cash (`Cash - Debt`), and `Net Cash / MCap`. | |
| | `current_ratio`, `liab_eq`, `fcf`, `profit_years` | `float`/`int` | Current assets / current liabilities, total liabilities / equity, free cash flow (`CFO - CapEx`), and profitable years out of the last 3 FYs (`0–3`). | |
| | `share_chg_3y`, `rev_growth`, `buyback_yield`, `div_yield`, `goodwill_pct`, `acq_pct` | `float` | 3-year share count change, YoY revenue growth, buybacks/MCap, dividends/MCap, goodwill/assets, and acquisitions/assets. | |
| | `period`, `fy`, `assets`, `ca`, `cl`, `cash`, `equity`, `goodwill`, `debt`, `debt_known`, `liab` | `mixed` | Raw SEC XBRL balance-sheet values (USD) and reporting period (`CY2026Q3I`, etc.). | |
| | `shares`, `shares_old`, `shares_old_period` | `float`/`string` | Latest common shares outstanding and shares outstanding 3 years prior. | |
| | `ni`, `ni_1`, `ni_2`, `oi`, `cfo`, `capex`, `buyback`, `dividends`, `acquisitions`, `rev`, `rev_1` | `float` | Raw SEC XBRL annual income statement and cash-flow values (USD) for the latest and prior fiscal years. | |
| | `insider_buy_value`, `insider_sell_value`, `insider_buyers_count` | `float`/`int` | Trailing 180-day open-market Form 4 purchase USD total, sale USD total, and distinct buyer count. | |
| | `tenk_date`, `tenk_accession`, `tenk_form`, `tenk_url` | `string` | Filing date, SEC accession number, form type (`10-K` / `20-F`), and direct SEC EDGAR archive URL. | |
| | `ai_verdict`, `ai_conviction`, `ai_balance_sheet`, `ai_value_trap_risk`, `ai_one_liner` | `mixed` | AI 10-K analyst summary fields (`CONTRARIAN BUY CANDIDATE` / `WATCH` / `AVOID`, conviction `0–100`, balance sheet rating, value-trap risk, and one-line thesis). | |
|
|
| ### 2. `ai_10k_memos` (21 Columns) |
|
|
| | Column | Type | Description | |
| | :--- | :--- | :--- | |
| | `ticker`, `name`, `cik`, `exchange`, `sector` | `string`/`int` | Company identifiers and SIC industry classification. | |
| | `accession`, `filing_date`, `read_on`, `filing_url` | `string` | SEC accession number, filing date, scan date analyzed, and direct EDGAR filing URL. | |
| | `verdict`, `conviction`, `balance_sheet`, `value_trap_risk` | `string`/`int` | Overall verdict, conviction (`0–100`), balance-sheet rating (`fortress`/`solid`/`stretched`/`weak`), and value-trap risk (`low`/`medium`/`high`). | |
| | `one_liner`, `business`, `why_hated`, `capital_allocation`, `management` | `string` | Concise investment synthesis covering the business model, market pessimism, buybacks/dividends/M&A, and leadership. | |
| | `bull`, `bear`, `red_flags` | `list` / `string` | Bullet lists of key bull points, bear points, and accounting/governance/footnote red flags from the 10-K. | |
|
|
| ### 3. `insider_trades` (11 Columns) |
| |
| | Column | Type | Description | |
| | :--- | :--- | :--- | |
| | `date` | `string` | Transaction date (`YYYY-MM-DD`). | |
| | `ticker`, `company_name`, `cik` | `string`/`int` | Issuer ticker, company name, and SEC CIK. | |
| | `owner`, `title` | `string` | Reporting insider name and relationship/officer title (`CEO`, `CFO`, `Director`, `10% owner`). | |
| | `kind` | `string` | Open-market transaction code: `P` (Purchase) or `S` (Sale). | |
| | `shares`, `price`, `value` | `float` | Number of shares traded, execution price per share (USD), and total dollar value (USD). | |
| | `form4_url` | `string` | Direct URL to the primary SEC Form 4 XML filing on EDGAR. | |
|
|
| --- |
|
|
| ## Data Integrity, Privacy & Caveats |
|
|
| - **Zero Credentials / Zero PII Guarantee:** Every export passes an automated security gate (`assertNoCredentials`) that enforces an explicit column allowlist, scrubs email addresses, and scans against secret/token patterns before writing or uploading. |
| - **XBRL Tagging Nuances:** SEC fundamentals reflect company-reported US-GAAP XBRL tags. Unusual tagging or dual-class share structures can occasionally distort market cap or debt ratios; always cross-check the linked SEC filing (`tenk_url`). |
| - **Research Only:** This dataset is for quantitative, academic, and machine-learning research only and never constitutes financial or investment advice. |
|
|
