Synth is a decentralized prediction engine built on the BitTensor network, designed to create a powerful forecasting tool for financial markets. The “product” is a system that incentivizes miners to build probabilistic models based on historical data using Monte Carlo simulations, producing a variety of potential price paths for assets like Bitcoin. These simulated paths model the dynamic price movement of assets, accounting for volatility, price clustering, long-tail risks (like Black Swan events), and sudden shifts in price.

Miners are incentivized based on how accurately their predictions align with actual future prices, but instead of a simple price target, they are assessed on how well they predict entire price paths over time. This process is made possible by a scoring system that rewards miners for providing probability distributions, not point predictions, allowing for a much richer understanding of future price behavior.

The system generates an impressive 165 million data points per day, compared to the 10,000 data points typical in other price prediction subnets. The platform’s architecture allows validators to send requests to miners, who respond by generating predictions in the form of 100 simulated paths of Bitcoin’s price over the next 24 hours. Every 24 hours, those paths are evaluated against the actual Bitcoin price path to see how accurate each miner’s prediction was.

Miners’ performance is rated based on a proper scoring rule, namely the Continuous Ranked Probability Score (CRPS). This scoring rule helps to avoid common pitfalls in other predictive models that may rely on overconfident predictions, thereby enabling a fairer and more robust way of rewarding the best models. Additionally, because the system models volatility as well as price, it captures the probability distributions that describe the most likely movements, making it incredibly useful for traders and investors in managing their portfolios.

• Miners (Forecast Generators): Each hour (approximately), registered miners receive a request to forecast an asset’s future price distribution. They must return N simulated price paths covering the specified horizon. In Phase 1, the typical task is to generate 100 Monte-Carlo price paths for Bitcoin over the next 24 hours at 5‑minute intervals. Importantly, miners forecast distributions, not single prices. Their simulations should capture realistic market dynamics (e.g. volatility clustering, fat tails) reflecting the miner’s belief in the probability distribution of future prices.
• Validators and Scoring: After the forecast period, validator nodes compare each miner’s ensemble of predicted prices to the actual market price path (using an oracle such as Pyth for real data). Accuracy is measured using the Continuous Ranked Probability Score (CRPS), which rewards both calibration and sharpness of the probabilistic forecast. A lower (better) CRPS means the forecast distribution closely matched reality. Validators compute the CRPS sum for each miner across all time increments. These raw scores are then normalized (via a softmax-like function) so that better forecasts receive higher normalized scores. Any miner who fails to submit valid forecasts is scored minimally (effectively zero) for that round.
• Leaderboard and Rewards: Synth maintains a long-term leaderboard for miners. After each round, the validator updates each miner’s leaderboard score by blending their recent scores with past performance, using an exponentially decaying time-weighted average (putting more weight on recent results). At the end of each day, these time-weighted scores are exponentiated (to amplify differences) and used to allocate the subnet’s token emissions proportionally. In practice, this means consistently accurate predictors earn a larger share of TAO tokens. This competitive incentive design encourages miners to continually improve their models and ensures that the most reliable forecasts are rewarded.
• API and Data Access: Synth provides APIs and tools so developers and agents can retrieve its forecasts and statistics. For example, the official Synth Subnet API (v1.0) exposes endpoints for retrieving the latest and historical prediction rates and the current leaderboard. Documentation is available for obtaining miners’ scores, validation results, and forecast data, making it straightforward to integrate Synth’s outputs into other systems. A live visualizer (on the Synth website) displays real-time forecast distributions, and historical synthetic data can be queried for backtesting or model training.

Technical Architecture

Synth Subnet architecture overview. In the Synth subnet, participants interact via two primary roles on the Bittensor blockchain: miners (who produce forecasts) and validators (who score forecasts). Miners submit their price-path predictions on-chain, validators fetch these predictions and compare them to live market prices (e.g. via a Pyth oracle), then compute CRPS scores and update rewards.

Synth is implemented as a Bittensor subnet (SN50) on the Bittensor network. As such, it inherits Bittensor’s distributed consensus (known as Yuma consensus) for ranking and rewarding contributions. Miner predictions are recorded on the subnet’s chain, and the CRPS scoring and emissions logic run according to Bittensor’s protocol. In effect, Synth’s intelligent forecasts become digital commodities on-chain. The subnet also provides client libraries and plugins, so on-chain agents (like Mode’s AI trading terminal) can query Synth forecasts directly. Mode has indicated that Synth’s live forecasts will be fed into its AI Terminal and trading agents for real-time risk assessment.

On the developer side, Synth’s GitHub docs outline the architecture: a loop where the validator requests predictions, miners respond with price-path arrays, and the validator runs a “prediction results” validation function before storing the data. Errors are logged and correct results are archived. Over time, the built-in scoring and leaderboard logic continuously evaluates each miner. Because the network is permissionless, any qualified model can participate. Mode Network has also created a dedicated discussion channel on the Bittensor Discord for Synth, fostering community participation.

Features and Use Cases

High-Fidelity Synthetic Data: Synth’s product is a rich probabilistic price dataset. By generating many possible future paths, Synth exposes AI models to scenarios (including extreme events) that may be absent in historical data. This synthetic data is intended for training and augmenting financial AI agents. For example, AI agents can consume Synth-generated paths during offline training to better learn market dynamics under uncertainty.

DeFi Agent Support: The subnet is explicitly designed to improve “forecasting and probabilistic reasoning of Agents and LLMs across decentralized finance”. In practice, Synth forecasts can be used by autonomous trading agents to make decisions. For instance, a DeFi agent might query Synth for the probability distribution of BTC prices over the next hour and adjust its trading or hedging strategy accordingly. Synth forecasts could also feed into on-chain risk protocols (e.g. predicting collateral liquidation risk).

Financial Applications: The synthetic distributions enable advanced financial modeling. Agents and analysts can use Synth’s data for option pricing (Monte Carlo simulations of price distributions), portfolio optimization, and risk management. The Synth website highlights specific use cases:

• Options Trading Platforms: Using Synth’s Monte Carlo paths to estimate fair option prices at different strikes.
• Perpetual DEXs & Lending Protocols: Forecasting liquidation probabilities and margin risk by evaluating the full distribution of future prices.
• Automated Market Makers (AMMs): Helping liquidity providers assess probability distributions of price ranges to optimize concentrated liquidity positions.

Institutional users can integrate Synth’s data into their own AI systems to enhance forecasting reliability, while retail traders can use the Synth interface (via LLM or dashboards) to receive actionable tips and probability-based insights.

Competitive Data Layer: Because data scientists and teams can submit their own models, Synth creates a marketplace of forecasting algorithms. Top performers (miners) see their models rewarded and their outputs become part of the public synthetic dataset. This competitive aspect aims to continuously refine the data quality. Mode’s blog emphasizes that Synth “incentivizes miners to continuously submit synthetic models which capture the full distribution of possible price movements” – effectively crowdsourcing the creation of high-quality synthetic data.