Table 1.
The TCFD disclosure categories and underlying recommended disclosures.
Source: [15].
Table 2.
Size and region of TCFD-supporting banks.
Table 3.
Sample composition.
Table 4.
Cleaning steps of parsed raw texts.
Fig 1.
On the left hand side, the BART model is pre-trained on all English Wikipedia articles and the BooksCorpus dataset. By masking parts of sentences ([MASK]), the model is trained to learn the semantics and to predict the missing parts. The process is repeated for all sentences in the pre-training dataset. On the right hand side, the model is fine-tuned on the MNLI task and returns probabilities for entailment, contradiction and neutral, as shown on the left hand side. Source: Own representation.
Table 5.
Overview of TCFD labels.
Fig 2.
Label evaluation matrix based on test dataset.
This matrix presents the results of the zero-shot text classification applied to our test dataset. The fine-grained TCFD labels are based on the recommended disclosures and the supplemental guidance for the financial sector.
Table 6.
Comparison of performance based on F1 scores.
Table 7.
Mean of label probabilities at category level per financial year.
Fig 3.
Climate-related disclosures by broad TCFD categories.
Fig 4.
Climate-related disclosures by fine-grained TCFD labels.
Table 8.
Mean of label probabilities for fine grained labels per financial year.
Table 9.
Mean differences in percentage points of climate-related disclosures.
Table 10.
Tukey difference-in-mean test of climate-related disclosures.