Danu Insight's mission is to change the understanding of systemic risks in financial markets, with an initial focus on the crucial yet underappreciated area of climate lobbying. We believe that legislation is the most important means to bring about the deep and rapid societal changes needed to achieve the goals of the Paris Agreement. The primary obstacle to enacting this legislation is corporate influence through lobbying activity, which often aims to weaken the ambition of legislation.
Institutional investors, who face systemic market risks from climate change due to their extensive market exposure, are best positioned to address this issue due to their influence. However, Danu recognised that there is currently no way for these institutional investors to track direct and indirect corporate influence across the entire market due to a lack of data. Using advanced data science and cutting-edge AI, we have developed systems that provide lobbying text analysis 50 times faster than manual methods. The prototype illustrates the application of Danu Insight’s Governance Scoring and Paris-Aligned Scoring to a universe of around 5000 publicly listed companies.
The “Companies” page provides a comprehensive list of all companies in our universe, along with their overall scores for both Transparency and Paris-Aligned Lobbying. The table can be sorted by these metrics, and users can search for specific companies using the search bar.
By selecting a company name, users can access detailed information across multiple tabs. The tabs at the top of the page offer insights into the company’s performance in areas such as Transparency and Paris Alignment.
Under the Transparency tab, users will find scores and commentary on key areas including Direct Lobbying Transparency, Trade Association Lobbying Transparency, Lobbying Governance, and Disclosure Inconsistencies.
The Paris Alignment tab offers a detailed breakdown of the company’s lobbying activities, categorized by various climate legislation topics. Each category features a score and an accompanying comment explaining the rationale behind the score. The overall score displayed on the main page is an average of the scores across these categories.
Our process works by scraping and aggregating a wide variety of raw data. This includes from government consultation responses from several key jurisdictions, having built automated data collection tools to download all the public responses in the US, Canada, UK, EU, Australia, and New Zealand. We have identified a total of 40 such repositories we need to write code to scrape.
We also track everything companies publish. We do this through bespoke code that relies on several open-source software packages that are effectively the same technology Google uses to aggregate all the copy on company websites.
We then process the plain text with our fine-tuned machine-learning models. We constantly track the best available technology for this. The fine-tuning process Involves us creating usually around 2,000 rows of examples of how the original data looks, instructions for how to approach each row, and perfect examples of how we want the text output and scoring to be.
We use these different models to achieve the following:
We separate the text into distinct paragraphs related to the same subject (e.g., a discussion on net zero). Each paragraph is categorized against a predetermined list of climate-related issues and non-climate text (e.g., financial information).
We then divide the text about climate change into small enough chunks for individual processing. For each chunk, we deploy four other finely tuned models to analyse the following using our numerical and qualitative methodology:
i. Transparency on engagement with climate policies: This includes whether they disclose the policies they engage with, the mechanism of engagement, and the policy outcome sought.
ii. Transparency on trade association lobbying: We assess whether they disclose their trade associations' climate lobbying positions, whether these positions are consistent with their own, and what actions, if any, they are taking to influence alignment.
iii. Climate lobbying governance: We evaluate whether they have disclosed processes to ensure alignment between their direct and indirect climate lobbying influence.
iv. Paris-aligned lobbying: This is the most complex model that assesses whether the objective of companies' communications with policy makers supports the aims of the Paris Agreement. The output from each model is a series of scores assigned to a chunk of text, along with a qualitative explanation of what the company is doing. We discard anything scored as NA (i.e. talking about a climate policy neutrally without intent to influence) simultaneously.
We use the assigned categories for each chunk of text to produce cohesive, analytical summaries for each category of climate legislation, again using AI. For example, all analyses of BP lobbying labelled "Carbon Tax" would be grouped and then condensed into a single overview of its lobbying on this issue.
It has taken us over a year and a lot of innovation to design and implement this process. However, all the models and steps require improvement. Wider data collection and more high-quality training data for our models will improve performance. In the future, we intend to apply a similar process to biodiversity lobbying, and the same process could theoretically be applied to any lobbying or systemically Important issue.