Bumper is a non-custodial, permissionless DeFi protocol which establishes a market for measuring and exchanging blockchain-based asset price risk. The protocol exposes a set of autonomous functions that allows users to take positions which protect the value of deposited cryptocurrencies such as Bitcoin (wBTC), Ether (wStETH), and others. The protocol locks these deposits for a fixed period. Protection ‘Taker’ deposits are pooled and collectively incur a regular, dynamic premium. Liquidity is supplied by Makers who deposit stablecoin (USDT) into a corresponding pool that serves to back the protected value of Maker assets. A novel rebalancing mechanism maintains an equilibrium of Taker and Maker pools should they diverge too far from their parameterised thresholds.
Since launching on Arbitrum in December 2023, Bumper has consistently outperformed traditional Put Option pricing on Deribit by as much as 30%. This is the first mechanism in 50 years with a statistically significant improvement over traditional Black-Scholes Option pricing.
Simultaneous with the smart contract development, Bumper spun up an AI Engineering team to radically enhance the protocol’s performance using a combination of three AI tech stacks to provide prediction, sentiment and technical analysis to the protocol. These inputs will be used to augment both the premium and rebalancing modules.
Bumper AI engineers are training a 70 billion parameter Large Language Model (LLM) with financial data consisting of Open Price, High Price, Low Price, Close Price and Volume metrics of Bitcoin price datasets. The LLM will be fine-tuned with Reinforcement Learning from Human Feedback techniques (RLHF) to reward the LLM for making predictions that agree with normalised actual price data. Bumper AI engineers started with training datasets of daily price and hourly data but are ultimately aiming to train on the LLM on tick data.
Since BTC tick data constitutes hundreds of terabytes of data, our engineers will reformat the datasets using Retrieval Augmented Generation (RAG), a relatively new approach to transforming information from relational database structures into Data Vectors. RAG technology will not only significantly enhance the performance of our LLM, but it is also expected to enable the seamless integration of multiple asset price feeds in real-time (a crucial requirement for RLHF) and operate within the constraints of existing LLM context windows.
Bumper uses a pre-trained Large Language Model to ingest vast amounts of financial Natural Language Processing (NLP) data to broadly categorise speculator disposition into a detailed sentiment score and distribution, giving a nuanced understanding of market mood. The LLM is fine-tuned utilising Bidirectional Encoder Representations from Transformers to label opinions, attitudes, and emotions, with particular NLP training to identify distinct financial vocabulary. Equipped with attention mechanisms and transformer-based structures, the LLM discerns market sentiment and investor behaviour to signal future market trends.
Bumper AI engineers have adopted a novel approach to train a Large Language And Vision Assistant (LLAVA) by extracting price data, and various indicators such as RSI and MACD, from an image, followed by labelling technical markers like support/resistance. This represents an end-to-end multimodal model that connects a vision encoder and LLM, which processes both price plots and NLP technical indicators, and trains the model on technical analysis using Long Short-Term Memory (LSTM) historical time series prediction.
Recent breakthroughs in LLM compute, LLAVA, RLHF, RAG, and LSTM are collectively making vastly improved predictions of market direction. Bumper’s triple-tier AI approach of combining pattern prediction, market sentiment and AI technical analysis is a powerful augmenter of the Bumper Protocol.
Bumper’s ability to dynamically price premiums according to real-time volatility, makes it an ideal candidate to ingest signals from LLMs to pre-empt market trends and rebalance proactively.
Analysis using Bumper’s proprietary Agent-Based Modelling (ABM) forecasts an economic improvement of 5-25% in terms of the protocol’s efficiency to balance the trilemma between Lower Premiums, Higher Yields and Solvency Robustness.
Disclaimer:
Any information provided on this website/publication is for general information purposes only, and does not constitute investment advice, financial advice, trading advice, recommendations, or any form of solicitation. No reliance can be placed on any information, content, or material stated on this website/publication. Accordingly, you must verify all information independently before utilising the Bumper protocol, and all decisions based on any information are your sole responsibility, and we shall have no liability for such decisions. Conduct your own due diligence and consult your financial advisor before making any investment decisions. Visit our website for full terms and conditions.
