Cracking the Code: Enhancing Development Finance Understanding with Artificial Intelligence

Every climate-COP reopens the same question: are rich-country pledges actually being delivered? The honest answer is surprisingly hard to produce, because the accounting system we use — OECD Rio markers — is a self-declaration regime. Donors decide which of their projects count as climate-relevant, and those decisions are not independently verified.

A project narrative — what the aid agency says the project does — is a richer, harder-to-game signal than a binary flag. This paper treats the narrative as the primary data, lets a clustering algorithm discover the latent topic structure of global aid, and then compares the resulting topic-based aggregates with the official ones. The divergence is the reporting gap.

The paper builds most directly on Toetzke, Banholzer & Feuerriegel (2022, Nature Communications), who used Doc2Vec + K-means on a subset of the CRS to recover ~180 activity clusters for climate finance. I extend that work in three directions: (i) the full CRS (5M projects, 1973–2022) rather than a climate-only subset; (ii) a modern embedding stack (multilingual MiniLM-L12-v2 + UMAP + HDBSCAN) that does not require specifying the number of clusters a priori; (iii) an LLM-assisted labelling layer (Zephyr-7B-β on top of c-TF-IDF) for human-readable topic names.

The climate-finance measurement literature has independently raised concerns about Rio-marker reliability — see Weikmans & Roberts (2019), Michaelowa & Michaelowa (2011), and Junghans & Harmeling (2012) — but these critiques have mostly relied on sampled audits of a few hundred projects. This study offers the first population-scale, fully automated, reproducible cross-check. It also complements supervised approaches such as ClimateFinanceBERT (Toetzke et al., 2023), which estimates climate finance via labelled training rather than unsupervised discovery.

1. Data

OECD Creditor Reporting System, ~5 million project records, 1973–2022. The text input is the concatenation of project title, short description, and long description. Declarations are multilingual; no manual translation is performed.

2. Embeddings

paraphrase-multilingual-MiniLM-L12-v2 (sentence-transformers) maps each narrative to a 384-dimensional dense vector. Multilingual by design — English, French, German, Dutch, Spanish, etc.

3. Dimensionality reduction

UMAP to 12 dimensions, preserving local and global structure while sidestepping curse-of-dimensionality issues that degrade Euclidean-based clustering.

4. Clustering

HDBSCAN (min. cluster size = 500) — density-based, hierarchical, optimal-k free, tolerant of the noise/"generic project" problem. Outliers are explicitly identified rather than force-assigned, then optionally re-absorbed via a second nearest-cluster pass.

5. Labelling

Class-based TF-IDF (c-TF-IDF) extracts the most discriminative tokens per cluster; the top-10 tokens and 5 representative narratives are passed to Zephyr-7B-β (fine-tuned Mistral-7B) to generate human-readable cluster labels.

6. Validation

Clustering quality tracked with Silhouette, Davies–Bouldin, and Calinski–Harabasz scores. Outlier reduction trades coherence for coverage and is reported transparently.

• Climate-finance reporting gap: $30–40B/year. Aggregated narrative-based climate flows are systematically lower than Rio-marker totals throughout the 2010–2022 period. The gap widens post-Paris Agreement, consistent with over-labelling incentives after the $100B pledge.
• 406 vs. 234. The narrative-based taxonomy recovers 1.7× more distinct topics than the OECD purpose-code system, with finer granularity in thematic areas (e.g. "Government & Civil Society — general" breaks into ~30 sub-clusters).
• Fashion cycles. Topics such as microfinance, coffee-farmer support, and Amazon Indigenous conservation display clear rise-and-fall dynamics that binary sector tags cannot capture.
• Niche visibility. Small but policy-relevant flows — great-ape conservation, demining, LGBTQ rights, reproductive health — become individually trackable for the first time.

• Not a causal identification. The paper documents a measurement gap; it does not identify the causal mechanism (strategic over-labelling vs. reporting inertia vs. genuine ambiguity). That's the natural next paper.
• Narrative quality varies. Reporting habits differ across agencies and over time; the multilingual embedding absorbs much of this, but not all. Results are robust to outlier handling; see Sec. 4.3 of the paper.
• Lower bound. Because HDBSCAN sets some ambiguous projects aside as outliers, the climate-finance estimate is best read as a lower bound. The truth lies between my narrative-based figure and the Rio-marker figure.

Issues, pull requests, and forks are welcome — especially translations of the labelling prompts and extensions to non-OECD sources (World Bank, AIIB, AfDB).

If you work on climate-finance accountability, ODA measurement, aid-transparency NLP, or development-economics methodology — I'd like to hear from you. I'm a PhD candidate at CERDI (Université Clermont Auvergne, CNRS, IRD) working on development and environmental economics; this paper is part of a broader agenda on measuring public-policy flows from text.

pierre.beaucoral@uca.fr
pierrebeaucoral.github.io
LinkedIn · GitHub · Google Scholar