Vinicio DeSola, Kevin Hanna, Pri Nonis

• https://drive.google.com/drive/folders/1rcRXZhb3JLY3A_kIO8gMk8jacRyR-Ik6?usp=sharing

• https://www.researchgate.net/publication/334974348_FinBERT_pre-trained_model_on_SEC_filings_for_financial_natural_language_tasks

Goal 1 FinBERT-Prime_128MSL-500K+512MSL-10K vs BERT

• Compare mask LM prediction accurracy on technical financial sentences
• Compare analogy on financial relationships

Goal 2 FinBERT-Prime_128MSL-500K vs FinBERT-Pre2K_128MSL-500K

• Compare mask LM prediction accuracy on financial news from 2019
• Compare analogy on financial relationship, measure shift in understanding : risk vs climate in 1999 vs 2019

Goal 3 FinBERT-Prime_128MSL-500K vs FinBERT-Prime_128MSK-500K+512MSL-10K

• Compare mask LM prediction accuracy on long financial sentences

Goal 4 FinBERT-Combo_128MSL-250K vs FinBERT-Prime_128MSL-500K+512MSL-10K

• Compare mask LM prediction accuracy on financial sentences : can we get same accuracy with less training by building on original BERT weights.

• Prime Pre-trained from scratch on 2017, 2018, 2019 SEC 10K dataset

• Pre2K Pre-traind from scratch on 1998, 1999 SEC 10K dataset

• Combo Pre-trained continued from original BERT on 2017, 2018, 2019 SEC 10K dataset