This episode of Novelty Search offers a deep dive into Synth, a Bit Tensor subnet revolutionizing financial forecasting by generating probabilistic price paths, moving beyond simple point predictions to provide a richer data source for traders, AI agents, and decentralized applications.
Preamble: A Skeptical Start to a Promising Solution
- The host, Const, initially expressed skepticism about another price prediction subnet, given the inherent difficulty of forecasting and the existence of established financial markets.
- However, the Synth team, with their background in statistical finance, presented a non-naive approach, leveraging Bit Tensor's unique capability to create markets for sophisticated predictive models.
- Const notes, "It's a very difficult problem... Predicting the future is very difficult, right?" highlighting the challenge Synth addresses.
- This segment sets the stage for exploring how Synth tackles this "incredibly difficult problem" by focusing on probabilistic paths rather than absolute price points.
- Strategic Implication: Investors should note Synth's differentiated approach, which focuses on modeling uncertainty and distributions, a potentially more robust strategy than direct price prediction in volatile markets.
Team and Vision: Building a Predictive Intelligence Engine
- James, from the Synth team, introduced the core members: Sam (quant modeling, algorithmic trading), Nathan (lead architect, software engineering), Stas (large distributed systems), David (quantum finance research), and Amber (strategy and ecosystem development). They are also working with GTV for Bit Tensor ecosystem insights.
- James articulated Synth's grand vision: "to create this predictive intelligence engine for finance and beyond."
- The initial focus is on forecasting Bitcoin by generating synthetic data paths, demonstrating the power of this approach before expanding to other assets and industries.
- Speaker Analysis: James presents a clear, ambitious vision, grounded by a team with diverse expertise in finance, AI, and software engineering.
The Forecasting Data Challenge: Beyond Point Predictions
- James addressed the common skepticism that if one could predict the future, they wouldn't share the data. Synth's aim isn't perfect prediction but building superior models of the future for retail traders, pro-traders, and especially AI agents.
- A key insight is the inadequacy of point predictions (single price targets) for AI, which require probabilistic models. Probabilistic models offer a range of potential outcomes and their likelihoods, crucial for sophisticated decision-making.
- Synth generates approximately 165 million data points daily, vastly more than typical point prediction subnets (around 10,000), by focusing on entire predictive paths.
- Actionable Insight: The shift towards probabilistic forecasting is critical for AI-driven trading. Researchers should explore how such rich datasets can enhance AI model training and decision-making in finance.
Synth Architecture: Generating Monte Carlo Price Paths
- Sam explained that Synth is a competition for quant research teams and data scientists to build generative models for time series data, starting with Bitcoin.
- Miners generate Monte Carlo paths: multiple simulated future price trajectories. These paths aim to capture nuances like volatility dynamics, clustering, heavy tails (rare, high-impact events), and skew.
- Monte Carlo paths are simulated sequences of future prices, each representing one possible evolution based on the miner's model.
- In Phase One (live since January), miners predict 100 paths for Bitcoin's price 24 hours into the future, at 5-minute intervals (289 prices per path).
- Validators request these paths hourly. Miners have two minutes to run their models and submit the paths. Validators store these and, 24 hours later, compare them to the actual Bitcoin price path to score the miners.
- Strategic Implication: The architecture encourages continuous model refinement by miners, creating a competitive environment for developing superior forecasting models.
The Scoring Mechanism: Continuous Ranked Probability Score (CRPS)
- Sam detailed the scoring, which uses Proper Scoring Rules. These rules mathematically incentivize miners to predict distributions that match their true belief about future uncertainty.
- Proper Scoring Rules are a class of functions used to evaluate the accuracy of probabilistic forecasts, where the forecaster is best off reporting their true beliefs.
- Synth employs the Continuous Ranked Probability Score (CRPS). The goal is to minimize CRPS. A perfect (zero) score means predicting a completely narrow distribution exactly on the real price.
- CRPS measures the difference between the predicted cumulative distribution function and the observed outcome. It penalizes both for being wrong and for being uncertain (too wide a distribution).
- Sam stated, "the optimum thing to do is to predict what you think the underlying distribution will be." This means miners must balance precision with the risk of being wrong, avoiding overly narrow or overly wide predictions.
- Scores are calculated for 5-minute, 30-minute, 3-hour, and 24-hour intervals to test models across different time horizons and capture phenomena like volatility clustering (periods where high or low volatility persists).
- A smoothing process (exponential weighted moving average with a 7-day lookback and 3.5-day half-life) identifies consistently performing miners. Rewards are then distributed using a softmax function.
- Actionable Insight: The CRPS mechanism is key to Synth's quality. Investors should understand that this incentivizes genuine probabilistic accuracy, not just lucky guesses, making the resulting data more reliable.
Performance and Validation: Evidence of Effective Forecasting
- Sam presented data showing the miner rewards distribution forming a Pareto curve, indicating that top miners are being disproportionately rewarded, a hallmark of effective incentive mechanisms in Bit Tensor. He noted, "personally I think we can make it steeper... we want to more heavily reward the best miners."
- Validators show high agreement (vtrust) on miner scores, suggesting the scoring is robust and consistently applied.
- Crucially, analysis over a month shows that the same miners consistently perform well, indicating they possess genuinely superior models, not just luck.
- Strategic Implication: Consistent top performance by certain miners validates the subnet's ability to identify and reward high-quality predictive models. This increases confidence in the data generated.
Miner Strategies: From Simple Models to Sophisticated Volatility and Tail Risk Modeling
- Const and Sam discussed miner strategies. A base miner might use Geometric Brownian Motion (GBM), a standard financial model assuming price changes are normally distributed with fixed parameters.
- Geometric Brownian Motion (GBM) is a stochastic process often used to model stock prices, where the logarithm of the randomly varying quantity follows a Brownian motion (random walk) with drift.
- Sophisticated miners improve by:
- Modeling dynamic volatility (sigma): Adjusting for time of day (e.g., lower volatility when US markets are closed), day of the week, and longer-term volatility regimes.
- Incorporating fat tails (high kurtosis): Accounting for rare, extreme price movements (Black Swan events) that normal distributions underestimate. Sam highlighted, "it's just that all of the best miners there are incorporating fat tails."
- The data shows top miners effectively model these changing volatility patterns and fat-tailed distributions.
- Actionable Insight: Researchers can learn from these emergent strategies. The focus on dynamic volatility and tail risk is where significant alpha can be found in crypto markets.
The Power of Monte Carlo Simulations: Versatile Use Cases
- Sam explained that Monte Carlo paths are highly versatile because any probability or path-dependent question can be answered by analyzing the ensemble of simulated paths.
- Synth offers live metrics pages for several use cases:
- Liquidation Probabilities: For leveraged traders to assess the risk of their positions being liquidated within specific timeframes (e.g., next 6, 12, 24 hours).
- Options Pricing: Synth's path-derived option prices are remarkably aligned with market prices (e.g., from Deribit), suggesting the models capture market-implied volatility and distributions. This can fuel trading bots or educational tools.
- Concentrated AMM Liquidity Provision: Helping Liquidity Providers (LPs) on platforms like Uniswap v3 to set optimal price ranges, estimate time in range, and understand potential impermanent loss.
- Impermanent Loss is a potential risk for liquidity providers in AMMs, where the value of their deposited assets diverges from the value of simply holding them, due to price changes.
- Speaker Analysis: Sam clearly articulates the practical applications, demonstrating how abstract probabilistic forecasts translate into tangible tools for DeFi users and traders.
Performance of Metrics: Outperforming Benchmarks
- Sam presented backtesting results for liquidation probabilities, showing Synth's forecasts significantly outperform standard GBM models in predicting the likelihood of 1% price movements.
- He mentioned, "we kind of massively outperform the basic benchmark model," indicating a quantifiable edge.
- For options, direct comparison with market prices on Deribit shows strong alignment, and the team is keen to build automated trading systems to definitively test Synth's edge.
- Strategic Implication: Demonstrable outperformance against standard models is a strong signal of value for investors and a rich area for researchers to explore the sources of this alpha.
Meta-Models and Synthetic Data Value
- Synth utilizes a meta-model (an aggregation of predictions, currently from the top 10 miners) to generate its public forecasts. This "wisdom of the crowd" approach can outperform individual models.
- The vast amount of synthetic data generated by the subnet was Synth's original concept. This data can be valuable for training AI agents, especially given the limited history of real BTC price data under consistent market regimes.
- Sam suggested a potential "flywheel effect": "It's producing so much data. You can look at what the best miners are doing... train your models on that data and you get progressively more and more good data which gets progressively better and better."
- Actionable Insight: The concept of a meta-model and the potential for a data flywheel are powerful. Researchers could investigate optimal meta-model construction and how synthetic data can accelerate AI development in finance.
Commercialization Strategy and Progress
- James outlined Synth's multi-pronged commercialization strategy:
- B2B: Selling API access to data marketplaces, funds, and exchanges; integrating dashboards and tooling.
- LLM Integration: Enabling Large Language Models (LLMs) to query Synth for probabilistic forecasts (e.g., "What's the probability Bitcoin will be over $110,000 tomorrow?").
- Business to AI (B2AI): Providing data for AI agent training and backtesting.
- Early partnerships include:
- Chainrisk: Integrating Synth for economic security risk management in lending protocols.
- Giza: Using Synth data for their stablecoin and other trading agents.
- Mode: Synth data integrated into Mode's AI terminal.
- New Data: Planned educational seminars for traditional finance.
- Strategic Implication: Diverse commercialization efforts indicate multiple pathways to monetization and adoption, reducing reliance on a single use case.
Roadmap: Expanding Assets and Verticals
- James and Sam detailed an ambitious roadmap:
- Q2 Current: ETH predictions are being added. Miners were given notice to adapt models.
- Future Assets: Focus on correlated paths (e.g., BTC and ETH in a single path). Potential TAO token price prediction. Commodities (gold, silver), then equities (S&P 500, individual stocks based on demand).
- New Verticals (Q3/Q4 onwards): Expanding beyond finance to energy pricing, weather, traffic, and health, leveraging the same Monte Carlo path generation and CRPS scoring.
- Long-Term (2026+): "Powering our AGI overlords with probabilistic models," becoming a foundational forecasting engine for AI across all industries.
- Actionable Insight: The expansion roadmap signals significant growth potential. Investors should monitor the successful addition of new asset classes and verticals as indicators of the platform's scalability and adaptability.
The Grand Vision: A Universal Forecasting Engine for AI
- The ultimate goal is for Synth to be the go-to data layer for AIs requiring probabilistic forecasts, whether for LLMs answering user queries or autonomous agents making real-world decisions.
- Sam mused on the adaptability of CRPS: "the scoring mechanism or some slight adaptation of the scoring mechanism can work well for that [binary outcomes like 'is it sunny?']".
- This vision extends to providing world models for robots and AI systems across numerous domains.
- Speaker Analysis: Sam's articulation of the grand vision is expansive, positioning Synth not just as a crypto tool but as a fundamental piece of future AI infrastructure.
Monetization, Bit Tensor, and the Value of Digital Commodities
- The discussion turned to how Synth's success could benefit the broader Bit Tensor ecosystem.
- Const emphasized subnets proving the value of their digital commodity. He suggested, "If you're able to extract... percentage points of better accuracy in predicting the price of Bitcoin... the sky's the limit on what the multiple is there for the digital commodity."
- Joseph (via Const's summary) views subnet tokens like Synth's (Alpha token) as representations of their unique commodity, acting as both a medium of exchange and a store of value, akin to "mini-Bitcoins."
- The Synth team envisions their token potentially being used by AIs to pay for data access in real-time.
- Strategic Implication: The long-term value of Synth's token is tied to the utility and demand for its predictive data. Success in commercialization and integration with AI systems will be key drivers.
Conclusion: The Significance of Proper Scoring in Prediction Markets
- Const lauded Synth's approach, particularly its use of Proper Scoring Rules like CRPS. This methodology directly addresses a core problem in previous prediction markets, where incentive structures could inadvertently reward overconfident or Martingale-style betting rather than genuine predictive accuracy. Synth's rigorous, statistically sound approach to incentivizing probabilistic forecasting marks a significant step forward.
This episode underscores Synth's sophisticated approach to financial forecasting, leveraging Bit Tensor to create a valuable data commodity. For Crypto AI investors and researchers, Synth offers a compelling case study in applied probabilistic modeling, with significant implications for AI-driven trading, risk management, and the future of decentralized intelligence.
