LIVE · TAO
TAO$— SUBNETS— VALIDATORS256
Bittensor intelligence updates
Home / Interviews/ Cascade (SN91)’s Road From Miner…
INTERVIEWS

Cascade (SN91)’s Road From Miner Competition to Commercial AI Forecasting

Explore how Cascade turns miner competitions into AI forecasting models, with Ephemeris bringing financial, retail, and energy forecasts to businesses.

Cascade (SN91)’s Road From Miner Competition to Commercial AI Forecasting

Cascade (SN91) is training AI models to forecast changes in financial markets, energy demand, retail sales, and other data recorded over time. The team is expanding its miner competitions while offering forecasts through its Ephemeris API.

On the TAO Pill Podcast, co-founder Chris Hames explained how Cascade (SN91) plans to turn its training progress into tools for traders and businesses.

Gordon Frayne interviews Cascade’s co-founder

Chris’ background combines a PhD in computational biology with experience building crypto trading strategies. A friend introduced him to Bittensor through SN8, and he later mined Precog, Synth (SN50), and early Sportstensor.

He noticed that several subnets produced forecasts, but none trained the underlying foundation models. Discussions with Chris Zacharia and approval from Const led to Cascade’s mainnet launch roughly two weeks later.

Cascade asks miners to build synthetic data generators, then tests whether their output improves forecasting models across different fields.

Chris noted that its first competition improved a 4 million parameter Datadog model using roughly 15% of the original training volume. It surpassed the original model and larger Salesforce models on key benchmarks, giving the team a basis for expanding the approach.

The second competition uses Cascade’s own mixture-of-experts model, with ~1.23 billion parameters and 120 million active during computation.

Planned expert lanes will target specific capabilities alongside the main competition. Chris also outlined a fine tuning competition using Mistral’s leading model. It will test forecasts across equities, crypto, and forex, with successful methods potentially supporting Cascade’s model and customer applications.

Cascade evaluates progress through a continually refreshed TS Benchboard, using fresh data and varied samples to limit overfitting.

Separate checks protect performance on external benchmarks. After ~150 billion training tokens, Chris placed the larger model within ~20% of Mistral’s leading model’s error.

The team expects further gains as training approaches a trillion tokens, though leading benchmark performance remains a target.

While training continues, Ephemeris combines forecasting models from providers including Mistral and Amazon into an ensemble designed to improve results.

The API supports agents and offers free sign-up credits. Cascade plans to add its own model and integrate Gamma. Commercial efforts include crypto trading platforms and a separately branded enterprise offering, with early discussions covering logistics, retail cash flow, and inventory forecasting.

Ephemeris by Cascade Website

Those applications give Cascade a way to test its value beyond benchmarks. Better forecasts can help businesses reduce stock shortages and excess inventory.

Chris said trading organizations had also seen Ephemeris outperform internal models in several scenarios. Forecasts of volatility, relative strength, and moving averages can support existing strategies, while customer feedback helps Cascade identify useful areas for specialized fine tuning.

Cascade’s Gnomon harness helps agents handle time series data without using future information and keeps records of their decisions. Chris reported ~15% profit in a month-long paper trading test using DeepSeek Flash, Hermes, and Gnomon on Hyperliquid.

The setup outperformed an unharnessed baseline. He also noted potential overlap with ChronoLLM (SN38)’s work on reducing look ahead bias in historical testing.

Miners need a Lium (SN51) account, and full model experiments can require eight H100 GPUs. Chris recommended testing ideas on smaller versions of the leading generator before scaling up. Each new winning generator is published on GitHub, where documentation also supports agent onboarding.

For 2027, Cascade’s priorities are a leading time series foundation model and external revenue that can fund token buybacks.

Chris emphasized consistent delivery as competition for investor attention grows, while welcoming Gamma’s potential to clarify compute and inference costs.

With training and commercial development moving together, Cascade’s next challenge is turning better forecasts into measurable value that keeps customers using its products.

⇥ Check Out Cascade (SN91) Here.

Enjoyed this article? Join our newsletter

Get the latest TAO & Bittensor news straight to your inbox.

We respect your privacy. Unsubscribe anytime.

The Daily Dispatch

Enjoyed this article?
Join our newsletter

Get the latest TAO & Bittensor news straight to your inbox — every morning before markets open.

IA
Ige A
Editor-in-Chief

No comments yet — be the first.

Leave a Reply