Longtail Token: Block Rewards for Evolution Strategies in Financial Machine Learning
Introduction
This document is the overview of a scientific experiment at the intersection of cryptocurrency technology, quantitative finance, and artificial intelligence. It is an ambitious endeavour by a Canadian Corporation Longtail Financial, lead by Founding Member Shawn Anderson, MSc. with a background in Big Data and AI. The company aims to bring to life, human-in-the-loop financial AI, in which a trained labour force works in tandem with AI technology to provide quantitative financial services, primarily in the cryptocurrency sector. The company is targeting the cryptocurrency sector because it is the sector that has the greatest demand for financial services, the most obvious of which being investment banking, wealth management, and trading. Speculation drives a vast portion of the cryptocurrency market. Longtail Financial is positioned to provide quantified financial services such as statistical portfolio optimization. Our technology has iterated and evolved over the past two years to bring you Longtail Token(LTT) and Foxhound, LTT being a bonding curve incentivization for distributed machine learning, and Foxhound being a robust, machine learning-driven financial engine that is the flagship software asset of Longtail Financial.
Longtail Token
The essence of this endeavour is to produce a revolutionary proof of work process that enables and incentivizes synchronous distributed training of machine learning market predictors. Specifically, we set up a proof of work algorithm, in which nodes are incentivized to produce greater performing market forecasting systems. We achieve collaboration through the use of standardized training sets, target sets, and performance metrics. TPOT, An open-source framework for genetic evolution of machine learning pipelines is used as a standard environment for predictor evolution. The result of the network is a continuously up-to-date set of high-performance market predicting machines. Market predictors can be used as general market forecasting oracles to aid in the holding, trading, and use of digital tokens.
Nodes that perform model training and model validation are rewarded with network tokens. Network tokens are bought back from the market by an automated fund that reflects the intelligence of the market predictors. As the fund’s performance increases, it will purchase more longtail token off of the market, thus increasing the price. Thus, we have a direct positive feedback loop from miner incentive to perform better and the value of their reward. Future research may apply tokenomics practices towards modelling the constraints of the network incentive ecosystem.
Model-driven proof of work environment gives an incentive to miners to perform the work better. In this case, nodes that produce high-yield models are incentivized by network tokens. Tokens are redeemable for a share of the network fund. Highest achieving market predictors are used by the network to manage a portfolio of decentralized assets. The greater the performance of the network, the higher the value of the token. The higher the value of the token, the higher the incentive for network performance.
Bonding curves may be used to create a quantified expectation growth curve based on an auto buying-selling mechanism which is common in current decentralized applications like uniswap, in which a fixed ratio liquidity pool is held between assets which allows for volume-proportional stability of price. Apply a bonding curve to that mechanism and you have the tools for a stable coin, or more interestingly, a quantified expectation growth curve.
Motivation
At longtail financial, we are consistently training AI algorithms to produce greater market predicting algorithms for use in proprietary trading. In April of 2020, we shifted research efforts towards and evolution strategies based machine learning library called TPOT. TPOT automates the role of the machine learning data scientist. It is used in our work as a proof-of-work algorithm. The objective of a worker node is to produce a model that is higher performing on a generalization test.
TPOT worked so well for us that we wanted to run bigger and bigger learning experiments with more data. The more data that we feed it, and the more training generations we allow it to run, the more sophisticated the predictors get.
But, the problem with TPOT is that it takes a long time to run because it runs a lot of simulations. You see, these simulations are very expensive. In fact, TPOT is deploying many generations of machine learning models. One generation at a time, it mutates a population, evaluates their performance, and then combines the higher performers into the next generation.
So I needed TPOT to synchronous its training across multiple computers. So I made a blockchain!
It’s an AI-driven blockchain, with humans in the loop. We are a team of engineers, designers, developers, and servers who want to bring this generation into the next generation of optimization for the planet. It’s through our optimization strategies that we can bring balance to the world. That’s why at longtail financial, we are looking particularly for industry partnerships in sustainability, energy technology, renewable materials, agriculture, as well as indigenous leadership.
You can read more about our investigation into sustainability and decarbonization in Shawn’s essay on financial land management for decarbonization.
A desire to scale lead us to two approaches 1. Using Amazon EC2 instances to massively parallelize the training process and 2. Synchronizing the training process across different machines. 3. Standardizing the data formats and performance metrics for training. With this infrastructure in place, we quickly realized that we had laid all of the groundwork for the implementation of a revolutionary proof of work algorithm to train highly useful machine learning models, in this case, predicting financial markets.
Technology
Standardized Financial Time Series
Firstly, a standard dataset is fetched. From here, a TPOT training process is run across all of the data available in a standardized training format.
Standard training data must have the following format:
[datetime, sector, broker, symbol, type, data…]
In this way, nodes can compile as much data as they desire to achieve higher model scores. A standard train set will be initially published, but it is expected that node operators will pursue deeper and more satisfying data sets to increase model performance. Additional features may be appended to the training algorithm for higher performance to be achieved. This will create a marketplace for private datasets to be marketed to supplement prediction algorithms.
Distributed TPOT training
This is how we run synchronous TPOT training across multiple machines. We use the term machine to represent a computing system with CPU and memory. For the most part, this has been our personal laptops and desktops, until we started using Amazon EC2. Distributed effort is achieved by standardizing train data formats, using the meta-programming of TPOT, using the validation techniques of machine learning, and using cryptographic methods of proof.
Model Selection
To start, we will have a block contain 10 models. Meta-data includes access information to the dataset that the model was trained on, and the scoring performance of the model. The score of a model is determined by how well it performs on generalized market forecasting, relevant market forecasting, and validation specific forecasting.
Distributed Model Validation
In the case of successful model training and submission, training data will be verified by validation nodes.
In the case that a node produces a better generalized model than is recorded on the public model record, the node may submit the untrained model to the network in the form of a pipeline. A pipeline is actually a python module. A python module is a text file that ends in the extension .py and can be interpreted by a python interpreter. The hash of a pipeline is used as the hash of the minted block. , and the pipeline can then be validated by validation nodes.
Specific, Relevant, and Generalized Performance Validation
Generalized market forecasting is evaluated by the model’s prediction capabilities on a random sample of all market data in existence. Relevant market forecasting is determined by a model’s ability to perform on recent market data in a single sector. Specific validation forecasting is determined by the validation performance of the dataset which it was trained.
Tokenomics
Nodes that produce high-yield models will are incentivized by network tokens. Tokens will are redeemable for a share of the network fund. Highest achieve market predictors are used by the network to manage a portfolio of decentralized assets. The greater the performance of the network, the higher the value of the token. The higher the value of the token, the higher the incentivization for network performance.
Tokens will be awarded to nodes that submit higher-performing prediction models. Tokens will be backed by an actively traded basket of digital tokens. The basket of digital tokens will be actively managed according to the predictions which are output by the currently highest-scoring market predictors. A node that has produced an active predictor will continue to be compensated as long as its predictor is used by the network to manage the network fund.
Tokenomics modelling software CADCAD will be used to model possible bounds on the price of LTFT. Inputs include, number of nodes, compute available, price of EC2 instances, dynamics of the market, advancement of AI techniques, token user speculation, predictor performance, agent performance, meta-agent performance. One might expect market predictors to perpetually increase in performance, however, given a crash in the value of LTFT, certain nodes might choose to stop contributing, resulting in market dynamics out-pacing the aggregate intelligence of the nodes.
Network Blocks
Blocks will be proposed when a complimentary model is discovered and proposed. A block is created that includes the new model. If the aggregate score of the block is higher than the score of the current HEAD block, then the longer chain will propagate and be adopted by the network. Nodes always adopt the longest chain. In the case of a tie, the chain with the higher aggregate score will be adopted and propagated.
Network Incentivizes
Nodes are incentivized to run TPOT training in pursuit of greater performant models. Models will be measured in performance as market forecasting machines. Nodes that find a greater scoring model will be incentivized with a digital token of value.
The Network Fund
The network fund will be actively managed by a set of reinforcement learning agents. Agents will perpetually output optimized portfolios given the environment produced by the active market predictors. A meta-agent will combine the portfolios of all active agents into an aggregate portfolio that will be deployed into a managed fund. At any point, LTT are redeemable for a portion of the managed fund.
Every ten minutes, the network fund will refresh its head block, and the current set of predictors will be used for environment creation. Nodes are incentivized for every head block that their submitted model is a part of. For the duration of the 10 minutes that the current HEAD block is selected, it’s predictions will be stamped on-chain, and used as a forecasted oracle of market behaviour. Each individual model in the HEAD block will make predictions for its relevant sector, broker, and asset basket.
By doing this, we are closing the loop on self-performing AI. We are unleashing the world’s compute power into orchestrated, objective-driven AI that will incentivize the perpetuation of itself. This will be the first step toward an intelligent blockchain.
Since the markets display novel emergence and metadynamics over time. Machine learning predictors will change in their performance over time, and, likely, there will always be a fleet of predictors coming online, and elder predictors falling offline. This enables an emerging meta economy of machine learning predictors to be the backbone of a truly AI-driven economy. So enough with the metaphysics, let’s get down to the implementation and testing.
Human-in-the-loop
Our approach to this endeavour is a human-in-the-loop mindset. We are young scientists and engineers that are openly witnesses to the incredible power of modern computing and connectivity. The future truly is faster than you think. We find that a simple mindset helps our team achieve very high results, and simultaneously appropriate prosperous relationships with the workforce of AI that has sprung up as fellow life on our planet. Machines are becoming more powerful than humans can imagine. LTT is here to make a play through that transition. The human-in-the-loop mindset is the idea that humans are powerful, and machines are powerful, and the most effective results can be achieved when humans and machines work synchronously in mass, so that is the primary objective that we are achieving. While growing our numbers, we are embracing extremely modern computing development processes such as web deployment, artificial intelligence, 3d design, and software architecture design. All synchronously resonating in an orchestral educational experience for all members of the organization.
FAQ
Why Blockchain?
It makes the training process more efficient, more stable, more simple, infinitely scalable, economically bootstrapped, and more hype!
What are the components?
Foxhound is the original LTT client, developed by LTF. But we expect that alternatives will emerge.
What does a block look like?
A block is a timestamped file containing the active market predictors and a hash to the previous block. A block is active while it is the HEAD block of the network. LTT follows Nakamoto consensus in which the longest chain is selected as the ground truth. LTT uses event-driven block times. Events being the improvement of the AI.
