English | 简体中文

RMRB Analysis

Overview

Analyze People’s Daily (人民日报, RMRB) advertisements by detecting ad pages, extracting ad blocks, running OCR, and generating LLM-based industry classification and commentary.

Workflow

Get original RMRB PDF
→ Detect advertisements by computer vision/text rules
→ Generate advertisement images (full/half/CV)
→ Extract ad blocks and filter duplicates
→ Convert images to text using OCR
→ Use AI to classify industry + ad type + region + analysis

Setup

1. Install dependencies:
   • pip install -r requirements.txt
2. Configure data paths in Config/Config.py:
   • MAIN_PATH / EXTERNAL_PATH / EXTERNAL_PATH_LIST
   • MODEL_PATH for PaddleOCR models
3. Download PaddleOCR model folders into MODEL_PATH (see PaddleOCR model list/releases):
   • https://github.com/PaddlePaddle/PaddleOCR/blob/main/doc/doc_en/models_list_en.md
   • https://github.com/PaddlePaddle/PaddleOCR/releases
   • Extract the archives so the listed folders are direct children of MODEL_PATH.
   • PP-DocLayout_plus-L_infer/
   • PP-OCRv5_server_det_infer/
   • PP-OCRv5_server_rec_infer/
   • PP-Chart2Table_infer/
4. Create Config/API.py (gitignored) with your LLM providers and keys:
   • Required by RMRBCore/RMRB_LLM_v5.py (MODEL dict with URL/Models/Keys)
   • Example:

     MODEL = {
       "GEMINI": {
         "URL": "",  # Not used for GEMINI; required for most other APIs
         "Models": ["gemini-2.5-flash"],
         "Keys": ["<api-key>"]
       }
     }

Data Layout

The analysis pipeline reads/writes under the data root (from Config/Config.py).

{DATA_ROOT}/
  {YEAR}/
    {YYYYMMDD}/
      {YYYYMMDD}.pdf or {YYYYMMDD}{VV}.pdf
  {YEAR}_AD/
    {YYYYMMDD}/
      {YYYYMMDD}_{VV}_FAD.png          # full-page ad
      {YYYYMMDD}_{VV}_HAD.png          # half-page ad
      {YYYYMMDD}_{VV}_CV.png           # CV-detected page
      {YYYYMMDD}_{VV}_{TYPE}_Block_{i}.png
      {image}.json                     # OCR + LLM outputs
    {YEAR}_Shape_Dict*.json            # shape filters & dedupe lists

Run Analysis (CLI)

Windows users can use the .bat scripts in Scripts/ (or PortableScripts/).

1. Prepare PDFs (optional tools)
   • python RMRB_PDFTools.py
   • Includes: MAC checker, folder formatter, existence check, PDF splitter, name fixer.
2. Generate ad images / blocks
   • python RMRB_AD_Image_Generator.py
   • Options:
     • Generate AD Image (FAD/HAD/CV)
     • Extract AD Block (creates *_Block_*.png)
     • AD Shape Analysis + Duplicate Check (writes final filter lists)
3. OCR
   • python RMRB_OCR.py
   • Writes OCR content into per-image JSON files.
4. LLM summary & classification
   • python RMRB_LLM.py
   • Generates Summary~... fields in the same JSON files.
5. Notebooks
   • RMRB_Analysis_v5.ipynb
   • RMRBCore/AD-Quant-Analysis.ipynb, AD_Text_Analysis.ipynb, etc.

RMRB Strategy

A research framework that asks one question of the advertisement archive: does what the People's Daily prints tell you anything tradable about Chinese equities?

The short answer, established below, is a little, but not enough. There is a statistically detectable contrarian signal in state-media attention, and it does not survive transaction costs. The framework is built to establish that kind of verdict honestly rather than to produce a flattering equity curve — most of the code below exists to catch the ways a backtest lies.

Modules

All under RMRBCore/, plain .py, importable independently.

Module	Role
RMRB_Dataset_v6.py	Raw ad JSON tree → daily industry panels. Filename parsing across both archive conventions, duplicate-creative screening, multi-model soft voting, page-area and prominence weighting, SW level-1 / level-2 tiers, CSV caching.
RMRB_Factor_v6.py	Share of Voice, seasonal detrending, trailing z-scores, activity masking, winsorising, cross-sectional standardisation, momentum/volatility controls, orthogonalisation, factor blending.
RMRB_Backtest_v6.py	Daily weight-drift portfolio engine with explicit execution lag, two-sided costs, turnover caps, overlapping tranches; APY / Sharpe / Sortino / Calmar / MDD / IR / alpha / beta / Newey-West t-stats.
RMRB_Research_v6.py	Rank IC, horizon scan, quantile ladder, event study, lead-lag, IS/OOS split, parameter grid, permutation test, deflated Sharpe, block bootstrap.
RMRB_Strategy_v6.py	Strategy registry — each entry states its economic hypothesis. Retains the original Daily_Z_Score_SoV.
RMRB_Pipeline_v6.py	CLI that runs the whole study and writes tables to Output/.

python RMRBCore/RMRB_Pipeline_v6.py --rebuild --figures

--level 2 switches the cross-section to SW level-2 sub-industries, --stage runs a single stage. The data root is found via RMRB_DATA_ROOT, then RMRBData/, then the drive list in Config/Config.py; the archive itself is gitignored.

What the framework adds to the notebooks

AD-Quant-Analysis.ipynb and Industry-Analysis.ipynb both begin from AD_Industry_Count_Analysis.csv, which reduces each ad to a hard count of its top-2 modal industries, commercial ads only. Four things get recovered here:

• Ad size and page position. A full page is not a half-page block, and page 2 is not page 16. Both are read from the archive (Shape_Dict pixel areas) and used as spend and prominence weights.
• Model disagreement. Each ad carries ~15 independent LLM classifications. Soft-voting keeps the confidence information a modal label throws away; mean agreement is 72% at level 1 and 58% at level 2.
• The policy channel. Public-service and policy ads were previously dropped. They are a different economic signal from a corporate media buy and get their own panel.
• Coverage. The legacy CSV totals 7,983 industry-counts; reproducing its own hard-count rule over the reconstructed 10,598 retained ads gives 10,466. The gap is mostly ads on the same page overwriting each other under a date-plus-page key.

Extraction: 10,598 ads (6,931 full-page, 1,436 half-page blocks, 2,231 CV-detected blocks), 143 dropped as duplicate creatives, 3,507 publication days from 2014-01 to 2025-12. Every retained ad maps to a valid Shenwan industry.

Method

Counts → Share of Voice → seasonal detrend → 12-month trailing z-score within each industry → activity mask → winsorise → cross-sectional rank.

Three of those steps are corrections to specific failure modes:

1. Within-industry z-scoring fixes the structural bias that sank the notebook's strategy. Banking is covered every month and cosmetics almost never, so a raw cross-sectional sort re-selects the same permanently-quiet sectors forever and becomes an accidental small-cap beta bet. Comparing each sector to its own history removes it.
2. Seasonal detrending is not cosmetic; see the trap below.
3. The activity mask drops sectors averaging under 0.5 ads/month over the trailing year, where a z-score just reflects whether one ad happened to run. About 24 of 31 sectors survive.

Backtests execute a month-end signal on the second trading day of the following month, let weights drift with returns between rebalances, and charge 15 bp one-way on both legs.

Findings

1. The contrarian signal is real but small. Attention today predicts weaker returns next month.

Panel	Rank IC	ICIR	t (Newey-West)
All ads	−0.0372	−0.187	−2.16
Prominence-weighted commercial	−0.0335	−0.141	−1.81
Commercial only	−0.0268	−0.116	−1.49
Area-weighted commercial	−0.0306	−0.126	−1.59
Public-service only	−0.0142	−0.050	−0.51
Legacy hard count	−0.0185	−0.090	−1.14

132 monthly cross-sections, ~24 industries each. The reconstructed panel roughly doubles the legacy count's t-statistic. Shuffling the score cross-section within each month gives p = 0.03, so the ranking is not an artefact of the score distribution. All 18 cells of the parameter grid carry the hypothesised sign.

2. The alpha window is one to three months. Contrarian IC by horizon: +0.037 (t = 2.16) at one month, +0.039 (t = 2.55) at two, +0.034 at three, zero by six, and negative by ten. Coverage marks a local top that unwinds over a quarter and then reverts.

3. The payoff is in avoidance, not selection. Quintile portfolios, % per month: Q1 (loudest) 0.314, Q2 0.459, Q3 0.653, Q4 0.635, Q5 (quietest) 0.659. The ladder is monotone (Spearman ρ = +0.90) but flat above Q3 — essentially all of the effect is the most-covered quintile underperforming. Owning the quietest sectors is not the trade; not owning the loudest ones is.

4. It does not survive transaction costs. The Q5−Q1 spread is 0.35%/month gross. A dollar-neutral book captures half of that and needs 3.2× annual turnover even with 3-month tranches:

One-way cost	0 bp	5 bp	10 bp	15 bp	20 bp	30 bp
Net APY	+0.99%	+0.67%	+0.34%	+0.02%	−0.30%	−0.94%

Break-even is around 15 bp, which is roughly the real cost of trading A-share sector ETFs. No long-only variant beats the equal-weight sector basket by a significant margin either: over the aligned 2015-01 to 2025-12 window the benchmark returns 4.01% APY, the best attention strategy (silent_winner, neglect blended with 1-month momentum) 4.74%, with an information ratio of 0.14 and t = 0.61. That is noise.

5. The macro timing signal is a seasonality trap. This is the most useful negative result. Aggregate commercial ad area, z-scored on a 36-month window, correlates −0.27 (t = −3.0) with next-month equal-weight sector returns, and an exposure overlay built on it returns 10.82% APY against the basket's 4.00% — +6.82% excess at t = 2.75. It is an illusion. RMRB ad volume is violently seasonal (December and March run four to five times February, one-way ANOVA p ≈ 1e-28) and A-share returns have their own calendar (January −3.3%, February +3.2%). Under a month-of-year control the correlation falls to −0.11 (t = −1.2) and the overlay's excess drops to 2.82% at t = 1.53. The apparent alpha was two calendars agreeing. macro_timing_exposure(deseason=True) is the default for this reason.

6. Breadth expansion fails. The Fundamental Law says IR scales with the square root of breadth, so moving from 31 sectors to 123 SW level-2 sub-industries should help. It does the opposite: at ~74 ads/month total, per-name flow becomes too thin, IC collapses to −0.011 (t = −0.79), the quintile ladder loses monotonicity (ρ = +0.30), and the sign flips between the in-sample and out-of-sample windows. Signal degradation outruns the breadth gain.

7. The signal is mostly independent of price momentum, and slightly stronger: neglect IC +0.037 (t = 2.16), neglect orthogonalised to 1-month momentum +0.028 (t = 1.56), momentum alone +0.023 (t = 0.92). The information is not a repackaged price factor — it is just not large enough to pay for its own turnover.

Out-of-sample. The split at 2021-01-01 was fixed before any strategy was sized. Contrarian IC is +0.043 (t = 2.11) in-sample and +0.031 (t = 1.02) out-of-sample: the same sign and a similar magnitude, with significance that does not survive the shorter window.

Verdict

State-media attention in the People's Daily is a genuine, measurable, contrarian sector signal with a one-to-three month horizon, concentrated in the underperformance of the most-covered sectors. Its effect size — roughly 0.35% per month on a top-versus-bottom quintile spread — is below the cost of harvesting it at any rebalancing frequency the signal decays fast enough to require.

It is usable as a risk overlay rather than an alpha source: hype_avoidance excludes recently-spiked sectors from an equal-weight basket at 0.9× annual turnover and tracks the benchmark within 0.13% APY, so the tilt is nearly free if you want the exposure reduction. Deploying it as a standalone strategy is not supported by this evidence, and the deflated Sharpe confirms the point: against 108 configurations tried, the observed Sharpe sits below what random search would be expected to produce.

Two caveats worth stating. The backtest trades Shenwan sector indices, which overstates achievable returns for sectors without a liquid ETF. And 132 monthly cross-sections is a short sample for a t-statistic of 2.2 — the finding is suggestive, not settled.

RMRB Online Downloader

CMD interface (RMRB_Online_v1):

(RMRB_Online_v1.1):

(RMRB_Downloader_v2) (Old name RMRB_Online)