Data Provenance

Sources

Complete transparency on academic foundations, data sources, and computation methodology for each dimension. Every number in the TGFI is traceable to its origin.

Verification note: All paper citations should be independently verified against the original source. DOI links are provided where available. Key findings are paraphrased, not direct quotes, unless marked with quotation marks.

Jump to:Trade Investment Technology Finance Leverage Policy

Trade

Bilateral trade flows, tariffs, and trade barriers

Weight: text 30% / hard 70%

Theoretical Foundation

Gravity with Gravitas: A Solution to the Border Puzzle

Anderson, J.E. & van Wincoop, E. (2003). American Economic Review, 93(1), 170–192.

DOI

Citation verified

Key Finding (paraphrased)

Bilateral trade flows are determined by relative trade barriers (multilateral resistance), not just bilateral barriers. A country’s trade with any single partner depends on its barriers vis-à-vis ALL trading partners.

How We Apply This

We use the gravity model to establish baseline expected trade levels for CN-US, CN-EU, and US-EU. Deviation of actual trade from gravity-predicted trade is our primary hard-data fragmentation signal: under-trading relative to the model indicates fragmentation; over-trading indicates cooperation.

Gravity Equations: Workhorse, Toolkit, and Cookbook

Head, K. & Mayer, T. (2014). Handbook of International Economics, Vol. 4, Chapter 3, 131–195.

DOI

Citation verified

Key Finding (paraphrased)

Provides the definitive estimation guide for gravity equations: PPML (Poisson Pseudo-Maximum Likelihood) with exporter-time and importer-time fixed effects, addressing zero trade flows, heteroskedasticity, and multilateral resistance.

How We Apply This

We follow their recommended PPML estimation methodology with proper fixed effects for our one-time gravity baseline calibration. This determines ‘expected’ bilateral trade volumes against which we measure deviations.

Shaping the World Economy: Suggestions for an International Economic Policy

Tinbergen, J. (1962). Twentieth Century Fund, New York.

Citation verified

Key Finding (paraphrased)

First formulation of the gravity model of trade: bilateral trade is proportional to the economic sizes of both countries and inversely proportional to the distance between them.

How We Apply This

Historical foundation. The gravity equation Trade_ij ∝ (GDP_i × GDP_j) / Distance_ij underpins our baseline model. Tinbergen’s original insight remains the core of modern trade modeling.

Data Sources

OECD Monthly International Trade Statistics

SDMX REST APIMonthly, ~60 day lag

Fields Used

Bilateral import/export values (USD millions)
By partner country (CHN, USA, EU27)
Monthly frequency, seasonally adjusted

Endpoint

https://sdmx.oecd.org/public/rest/data/OECD.SDD.TPS,DSD_ITF@DF_ITF_GOODS_M/

Sample API response

// Example: US imports from China, Jan 2026
{
  "structure": { "dimensions": { "series": [
    { "id": "REF_AREA", "values": [{ "id": "USA" }] },
    { "id": "COUNTERPART_AREA", "values": [{ "id": "CHN" }] },
    { "id": "FLOW", "values": [{ "id": "M", "name": "Imports" }] }
  ]}},
  "dataSets": [{
    "series": { "0:0:0": {
      "observations": {
        "2026-01": [35821.4],
        "2026-02": [33105.7],
        "2026-03": [36290.1]
      }
    }}
  }]
}

WTO Tariff Download Facility

Bulk download (CSV) + Tariff Analysis OnlineAnnual, ~6 month lag

Fields Used

MFN applied tariff rates (simple & weighted average)
Bound tariff rates
Bilateral preferential rates where applicable
By HS product code (6-digit)

Endpoint

https://tariffdata.wto.org/

Sample API response

// Example: US weighted average applied tariff on Chinese goods
reporter,partner,year,product,duty_type,avg_rate
USA,CHN,2025,TOTAL,MFN,19.3
USA,CHN,2025,TOTAL,Applied,24.8
EU27,CHN,2025,TOTAL,MFN,5.1
EU27,CHN,2025,TOTAL,Applied,8.2

UN COMTRADE

REST API v1Annual (final) & monthly (preliminary), variable lag

Fields Used

Commodity-level bilateral trade (HS 6-digit)
Trade value (USD) and net weight (kg)
All reporter-partner combinations

Endpoint

https://comtradeapi.un.org/data/v1/get/C/A/HS

Computation Pipeline

1Fetch bilateral trade flows

Pull monthly bilateral trade data (imports + exports) from OECD SDMX API for three pairs: CN–US, CN–EU, US–EU. Convert to USD at market exchange rates.

2Compute Trade Share Deviation (50% of hard data score)

S_ij = Trade_ij / Total_Trade_i

ΔS = (S_current – S_baseline) / S_baseline × 100

CN–US example: If US trade with China = $362B and US total trade = $2,140B, then S = 16.9%. Baseline (2019 avg) = 19.2%. ΔS = (16.9 – 19.2) / 19.2 = –12.0% → normalized to –38.0

3Compute Tariff Change Signal (30% of hard data score)

ΔTariff = –(applied_rate_current – applied_rate_previous)

Negative sign: tariff increase = conflict signal

CN–US example: If weighted avg applied tariff rose from 19.3% to 24.8%, ΔTariff = –(24.8 – 19.3) = –5.5 → normalized to –42.0

4Compute Trade Balance Trend (20% of hard data score)

ΔBalance = (balance_t – balance_{t-1}) / |balance_{t-1}|

Widening deficit with adversarial partner = conflict signal

Tracks directional change in bilateral trade surplus/deficit

5Blend Hard Data sub-components

HardScore = 0.5 × TradeShareDev + 0.3 × TariffSignal + 0.2 × BalanceTrend

All sub-components normalized to [–100, +100] via min-max on historical range

HardScore = 0.5(–38.0) + 0.3(–42.0) + 0.2(–15.0) = –34.6

6Compute Text Score from T1 academic sources

TextScore = mean(LLM_classify(doc_i)) for all docs in period

Sources: NBER working papers, IMF staff discussion notes, BIS quarterly reviews, OECD policy papers, central bank publications. Each document classified on [–100, +100] cooperation–conflict spectrum.

If 18 relevant publications this quarter, each scored by LLM, mean score = –51.3

7Blend to Trade Composite

Trade_Composite = HardScore × 0.7 + TextScore × 0.3

70% hard data weight because bilateral trade has reliable, timely quantitative data. 30% text captures policy signals and sentiment not yet reflected in flows.

Trade_Composite = (–34.6)(0.7) + (–51.3)(0.3) = –24.2 + –15.4 = –39.6

Validation

Text\u2013Hard Convergence

Pearson correlation between text score and hard data score across periods. Target: r > 0.6. If text and hard data tell conflicting stories, the convergence metric flags it.

Event Alignment

Score spikes should align with known trade events: tariff announcements, FTA signings, trade war escalations. Manual review each quarter against event timeline.

Gravity Residual Backtesting

Does the trade share deviation predict next-quarter actual trade volume changes? Granger causality test on historical data (2015\u20132025).

Weight Sensitivity

Monte Carlo: perturb text/hard weights by \u00b120% (1000 draws). Verify composite score ranking is stable across perturbations (rank correlation > 0.95).