harb/onchain/hAIrberger.md

The hAIrberger Protocol

Problem / Current Situation:

A static liquidity provider strategy in a dynamic market leads to:

• impermanent loss
• reduced earnings

Harbergs baseline-like liquidity setup can reduce impermanent loss only at the cost of liquidity share and fee earnings. Token-printing-priviliges gives unfair advantage, but not forever.

Succesfull/Dynamic LP strategies use indicators:

onchain:

• current and historic price
• current and historic volume
• current and historic liquidity distribution

offchain:

• TA
• Macro
• Sentiment

Making the Harberg LP strategy dynamic

• Oracle Problem #1: The offchain indicators can not be used by decentralized communities at all, having a unsolved oracle trust issue.

  • use staking, an egoistic marketplace, as data source for sentiment
    • % staked
    • average tax rate => egoistic oracle marketplace

• Using historic data: The onchain indicators need to be digested, e.g. TA to get signals.

  • algorithmic trading / backtesting => model => deep learning
  • fold historic data into AI model training

• Oracle Problem #2: Any algorithm running offchain can not be trusted by community. AI models can not run onchain. Trained models can't react to real-time data (see LMM knowledge cutoff).

  • use on-chain algorithm that works with real-time data, historic knowledge and limited gas => follow-up problem: "write an algorithm that uses limited gas to maximize fees earned, starting with no training data."

• Metaheuristic: is a higher-level procedure designed to find, generate, and tune partial search algorithm that may provide a sufficiently good solution to a machine learning problem, especially with incomplete or imperfect information or limited computation capacity.

  • Genetic algorithms are used to generate solutions to optimization problems via biologically inspired operators such as selection, crossover, and mutation.
  • selection can be applied on:
    • gas usage
    • fees earned after running scenarios
      • scenarios can be replaced by historic trading data after launch of project
  • result of evolution (latest generation) can be deployed onchain and integrated into recenter
    • recenter() => getDynamicLiqParams() => _scrapePositions() => _setPositions()
    • function getDynamicLiqParams(inputs) returns (params) is an EA living in upgradeable contract => follow-up problem: EVM op-codes can not be used in Evolution, too much grammar => follow-up problem: upgradeable contract needs governance

• Evolving Stack Machines instead of reverting, invalid op-codes should be ignored.

  • Push3 by Lee Spector => implement Push3 runtime in solidity

• Governance tbd

Conclusion

• a dynamic strategy is possible, first using human-generated solutions (sentiment function), and later upgrading into GA.
• basic governance to be solved before GA.