gm
SN28On-chain oracle network providing external data to the Bittensor ecosystem
Bittensor's S&P 500 Oracle, built by Foundry Digital. Miners compete to predict the index price every five minutes using open-source neural networks, with all models and training data published to HuggingFace.
// S&P 500 predictions, decentralized
gm (SN28) is Bittensor's dedicated financial forecasting subnet, operated by Foundry Digital, one of the larger institutional participants in the Bittensor ecosystem. On-chain the subnet goes by "gm," but the product is the Foundry S&P 500 Oracle: a decentralized competition where miners build machine learning models to predict short-term S&P 500 price movements, with every model and dataset required to be published openly on HuggingFace.
The simple version: It's like a public quant research competition where participants submit their stock prediction models openly, and the most accurate models earn emissions.
Centralized equivalent: Think Bloomberg's quantitative analytics or QuantConnect, but with fully open-source model transparency and decentralized incentives.
How it works:
- Miners build neural network models to predict the S&P 500 closing price for six five-minute intervals into the future. All models and input data must be open-sourced on HuggingFace to receive emissions.
- Validators send miners a future timestamp and collect their price predictions. Once the interval matures, validators score each prediction against the actual S&P 500 price using two metrics: Directional Accuracy and Mean Absolute Error. Miners are ranked relative to each other, creating a continuously competitive environment.
- The problem it solves: High-quality financial market prediction models are expensive, proprietary, and concentrated at a handful of quantitative firms. Individual traders and researchers have little access to transparent, open-source forecasting tools.
- The opportunity: A continuously running, open-source prediction market for the world's most-watched equity index, where competition keeps accuracy improving and the best models are publicly available.
- The Bittensor advantage: Market-aligned incentives. Miners only earn based on accuracy, and the requirement to open-source all models on HuggingFace keeps the competition transparent and harder to game. The adversarial environment pushes miners toward differentiated architectures and datasets.
- Traction signals: The subnet has been running since February 2024. On-chain data shows zero active miners at the time of this snapshot, which may reflect market hours timing since the S&P 500 is only active Monday to Friday during U.S. trading hours. Developer activity on the official Foundry Digital repo is maintained and includes an active CI pipeline.
Category: Financial Forecasting and Trading Signals | Centralized Competitor: Bloomberg Terminal, QuantConnect, Two Sigma (proprietary)
Financial market prediction is one of the oldest and most intensely competitive fields in computing. Access to high-quality predictive models has historically been limited to institutions with the resources to build and maintain proprietary systems. Foundry Digital, which operates one of the larger Bittensor validators, launched SN28 in February 2024 to decentralize that access.
Mechanism:
Miners on SN28 build neural network models that predict S&P 500 price movement over short horizons: the next six five-minute candles. A key design constraint is that all models and training data must be published openly to HuggingFace before miners can receive emissions. This prevents black-box gaming and creates an evolving public library of forecasting models.
Validators drive the competition by sending miners a future timestamp and collecting their price predictions. Once the prediction window matures, validators compare outputs against the actual S&P 500 price using Directional Accuracy and Mean Absolute Error. The external source of truth is Yahoo Finance's S&P 500 price, which makes the rewards mechanism objective and resistant to manipulation.
Price is up roughly 44% over the past 30 days and about 12% over the past seven days, with net positive inflows over the seven-day window. Pool depth sits around 14,900 TAO with a root proportion of approximately 0.16, indicating most of the pool comes from organic staking demand rather than protocol subsidy. Volume over the past 24 hours was around 3,260 TAO.
One data point that warrants monitoring: the snapshot shows zero active miners with 100% of the miner emission allocation being burned. The subnet's operation is tied to S&P 500 trading hours, so this may reflect an off-hours snapshot. Foundry's documentation explicitly warns that the base miner provided in the repo is not intended for mainnet, meaning competitive participants need to build their own custom models, which raises the bar for entry.
- Active miner gap: Current snapshot shows zero active miners with 100% of the miner emissions burned. The subnet only generates meaningful prediction activity during U.S. market hours. Sustained absence of miners during live market hours would signal real ecosystem attrition and is worth monitoring.
- Market hours dependency: Unlike most Bittensor subnets that run continuously, SN28 is fundamentally tied to U.S. equity market hours (Monday to Friday, 9:30am to 4:00pm ET). This creates predictable operational dead zones and may limit miner participation.
- High entry bar for miners: Foundry explicitly says the base miner will not earn on mainnet. Competitive participation requires building custom neural network models and publishing them to HuggingFace. This keeps noise out but also limits the miner pool.
- Thin on-chain identity: The on-chain identity fields for gm are unpopulated: no website, GitHub, or Discord linked. New participants have to find Foundry Digital's GitHub independently, creating discovery friction.