coindesk.com
At Spectral, we’re creating and incentivizing a network of modelers, creators, users and validators using proof-of-stake mechanics. The idea — similar to Chainlink (LINK) and The Graph’s (GRT) models — is to build a decentralized marketplace with a built-in feedback mechanism that ferrets out and discourages bad actors.
This article is part of CoinDesk's "Staking Week." James McGirk is a senior writer at Spectral Finance and the co-founder of Lonely ROCKS.
Our multi-asset credit risk oracle (MACRO) score is a machine learning model that weighs approximately 100 on-chain signals to produce a three-digit score predicting a wallet's likelihood of liquidation on an on-chain loan.The score is similar to the FICO score, and ranges from 300 (representing a very high risk of liquidation) to 850, representing a very low risk. It’s very similar to what you’d get from a traditional credit report, only instead of relying on Experian, Transunion and Equifax to keep tabs on your spending, you opt-in with your wallet.
The promise of an on-chain credit score is that it’s opt-in, completely transparent, and eventually, production of the algorithm generating the scores can be decentralized by incentivizing a competitive marketplace. Netflix pioneered the technique in the 2000s when they offered and eventually paid a million-dollar bounty to a team of data scientists who improved their recommendation algorithm by 10%.
To create a similar dynamic using smart contracts, we’re building a validator network. The traditional model is to pay rewards to a validator node for producing blocks and validating rewards, and punishing nodes — which is called slashing — by taking away their stake when they misbehave, which entail failing to maintain the node, behaving maliciously or other blockchain malfeasance.
Our model is a little more complicated. We’re incentivizing a contest. So we divide our network into modelers (who are machine learning engineers earning bounties by creating accurate models) and creators, who create data science challenges for the modelers to tackle, in this case an accurate credit score generated from on-chain information.
We also have validators, who vet the models for quality, and, after the contest ends, we have users who pay to use scores (i.e. machine learning inferences) generated from the winning models.
Creators post a bounty which modelers compete for during a specific time period; meanwhile, during this time modelers stake SPEC tokens, which can be slashed by validators in case of bad behavior. Creators post a service level agreement (SLA), which is standard legalese for contractual terms, which modelers commit to when entering the contest. This SLA can specify criteria such as achieving accuracy benchmarks or uptime.
See also: Staking Risks Are Vastly Misunderstood | Opinion
During the contest, validators police the terms of the SLA, slashing modelers who don’t adhere to them. Much like a blockchain internal affairs bureau, validators themselves can be slashed if they collude with modelers or don’t hit deadlines.
After the contest window closes, the consumption window begins, and users start using the model to produce credit scores from the winning model. Creators and modelers share earnings generated during this time and hopefully forever after. In essence, the idea is to use crypto to nourish a flourishing ecosystem that grows extremely accurate machine learning models as a byproduct.
Cryptoeconomics, when it works, creates a hothouse environment where ideas are iterated upon by people all over the world. Creditworthiness assessment is just one use case, by building on a blockchain, smart contracts can build off-chain processing (such as zero-knowledge machine learning) onto the system, so nearly any data set can be encrypted and worked on given enough processing power and time — whether it’s tumor hunting, medical records inferences, insurance payouts, bail calculations even training robotic operating systems to serve hamburgers.