We're a startup building end-to-end deep learning models with continual learning for autonomous trading on the stock market.

The most recent development model Sophon-2 is currently trading live on Nasdaq, with an annual return of +28%.

Our goal is to make these models widely available to retail investors.

Thank you for signing up.

Something went wrong while submitting.

about sophon

Sophon market performance

Sophon-2 is our latest model. Compared to Sophon-1, it handles changing market conditions in a better and more stable way. Sophon-2 trades stocks on Nasdaq-100 index, outperforming it by +50%.

Sophon-1 was our first technical trial in real market conditions, employing deep reinforcement learning and continual learning. It achieved a accuracy of 64.6% on long trades. However, as a small model, Sophon-1 struggled to adapt to changing market conditions and was retired at the end of 2023.

Compared to robo-advisors and hedge funds, Sophon gives consistently high yearly returns to low fees.

Robo advisors

Such as

Betterment, Wealthfront, Lysa, Opti

Invests in

Indexes, bonds
‍

Pros

Low risk

Low fees

Widely available

Cons

Low returns

Hedge funds

Such as

Renaissance, Two Sigma, D.E. Shaw

Invests in

Wide range of financial instruments

Pros

High risk

Widely available

Cons

High fees

Barries to entry

End-to-end AI

Such as

Sophon

Invests in

Stocks
‍

Pros

High risk

No % fees

Cons

New technology

Sophon-2 metrics

Current portfolio

Trades

This presents both historical and current trades on the US markets from the internal perspective of the AI, Sophon.

Market trade order history

Executed on alpaca.markets API.

Market trade order history

Executed on alpaca.markets API.

Our AI principles

Sophon is not a trading algorithm.

Instead of humans doing quantitative analysis, Sophon continuously explores and adapts to the market using end-to-end generalized AI models.

Principles

No train / test period

Continuous training from data start, no specific train/test periods.

Continual learning

All deep learning models undergo constant and frequent training.

No human labeling

The AI determines what is beneficial, with top-level behaviors rewarded accordingly.

Hyper parameters

No fine tuning or selection of hyper parameters. All parameters are used all the time.

1. Trading

Deep reinforcement learning for trading

We use it to trade: Agent takes in raw trading data and outputs raw trading actions.

2. Portfolio

Transformers for building portfolios

Transformers are used to make sense of long sequences are used in the stage where a portfolio are built up from raw trading actions. A complete agent is formed from this stage.

3. Final trading output

Ensembles of agents, meta-AI for final output

Agents, which utilize reinforcement learning for trading and transformers for portfolio construction, are generated from a large grid of hyperparameters. A meta-AI determines which agents to deploy in real-time, rather than retrospectively. Ineffective agents are gradually eliminated over time.

Roadmap

Timeline

2024

Sophon-3 development start

2024

Sophon-2

Sophon-2, launched in June 2024, features a significantly larger and more refined system of deep learning models compared to Sophon-1, utilizing a multi-layered transformer approach. Improved backtesting and simulations enhance its continuous evolution and training.

2022

Sophon-1

Sophon-1 is live trading in March 2022.

2018

Sophon-1 development start

Research and development began in 2018, exploring the use of continual learning and deep learning for trading through experiments and conceptual advancements.

the team

The founding team brings extensive experience in building successful startups, developing complex software, working in fintech and banking/funds, and advancing AI technology.

CEO & Founder

Hans Dahlström

Hans began programming at the age of 7 and, at 17, co-founded his first company, Starbreeze, which went public in 2001. In 2005, he left the game industry to co-found the B2B SaaS company Hansoft with two other partners. Hansoft was sold to Summit Partners in a significant exit in 2017, and Favro was spun out as a new startup. In 2018, Hans started studying advanced deep learning, including deep reinforcement learning and financial machine learning, alongside his work at Favro. This research led to a six-year project that culminated in the formation of SneakPeek, which entered startup formation in 2024.

hans.dahlstrom@sneakpeek.ai

Founder

Adam Tittenberger

Adam has studied economics, statistics, and mathematics at both Stockholm University and Uppsala University. He has over 17 years of experience working at various financial institutions, including mid-sized investment firms, banks, and fintech companies. Throughout his career, Adam has held diverse roles, such as data management, portfolio risk and performance analyst, AMA operational risk analyst, credit analyst focusing on Basel/IFRS9 topics, and conducting analytical work.

adam.tittenberger@sneakpeek.ai

Founder

Johan Baer

Johan has extensive experience in product design, art direction, and interactive design. He has worked with various companies, including the international consultancy firm Bontouch as a Senior UI Designer, the B2B SaaS company Favro as a Lead Designer, and his own company.

johan.baer@sneakpeek.ai

Founder

Karl-Oskar Lundin

Karl began programming at the age of 8. He studied computer science, economics, and mathematics at Uppsala University. Since 2007, he has successfully run a company, specializing in B2B, B2C, and SaaS. He has been awarded the Innovator of the Year by Dalarna Business. Karl possesses extensive knowledge in payment solutions and compliance regulations, including tax compliance.

karl.oskar.lundin@sneakpeek.ai