Late last year, I created several markets to assist with the development of my deep learning stock trading model. Thanks in part to Manifold machine learning experts, we have gotten it into live testing. However, as every machine learning engineer knows, and how long it took OpenAI to get the GPT store online shows, the infrastructure surrounding the models (like the API calls, the server hardware, and the traditional code that obtains the data) is far harder than training intelligent models.

Full details of the Seregon AI Investing organization: https://shoemakervillage.org/temp/transition_trio.pdf, pages 6-7.

The model is a binary classifier that uses 15+ layers of different types with custom activation functions. It started out as a cryptocurrency trading bot, but we transitioned to stock trading due to the lack of SIPC insurance at cryptocurrency exchanges. It is the best of 1400 models that were trained with different hyperparameters and layer configurations and data processing on a bank of 4090 cards over the course of a year. It was trained on more than 5TB of data related to financial markets from multiple exchanges.

The model has an accuracy of 92% and backtests at 656% CAGR on forward unseen test data when a fee/slippage of 0.15% on each buy and sell is assumed. It exclusively trades S&P 500 stocks. On January 22, 2024, its accuracy was 87% and it exceeded the backtesting (1.19% daily) by earning 1.31% at the time the exchange integration code crashed and failed to sell.

Earnings for January 22, 2024:

ID Pair Profit (USD)

----- -------- --------------

4 ANSS/USD -0.17% (-0.58)

23 DAL/USD 3.82% (12.78)

3 NWL/USD 3.95% (13.53)

1 BXP/USD 2.30% (7.95)

2 DFS/USD -1.24% (-3.75)

5 PVH/USD -1.05% (-2.54)

7 APA/USD 0.39% (1.32)

6 ARE/USD -0.88% (-2.18)

8 VFC/USD 4.39% (14.91)

29 USB/USD 1.13% (3.76)

27 ALGN/USD 0.35% (0.95)

24 VNO/USD 2.58% (8.64)

21 INTC/USD -1.47% (-4.97)

22 CFG/USD 1.22% (4.00)

28 UAL/USD 9.37% (32.40)

25 BA/USD 1.03% (2.19)

10 CTLT/USD 0.04% (0.12)

14 ALB/USD 6.09% (14.48)

20 PAYC/USD 2.37% (4.60)

16 ZBRA/USD 0.02% (0.04)

11 FCX/USD 2.16% (7.29)

12 CMA/USD 1.94% (6.18)

19 AMD/USD -1.26% (-4.26)

13 EXR/USD 0.20% (0.60)

17 CCL/USD 1.58% (5.46)

9 NCLH/USD 2.36% (8.20)

15 ALK/USD 4.63% (14.58)

18 KEY/USD 1.70% (5.76)

26 UAA/USD 2.22% (7.65)

----- -------- ------- --------------

Total 159.11 USD

When the infrastructure bugs are worked out and the full amount of real money is used, I will add a comment stating the model has started running in production. From that point, the total returns will be calculated.

If this model, or an improved derivative, doubles its starting capital within one year of that date, the market will resolve to YES immediately on the day the money is doubled. Otherwise, it will resolve to NO, including in cases where some unrelated issue, such as troubles with "traditional" software engineering and exchange integration prevents us from running the model for long enough to earn that much money.

The money earned will be calculated after all expenses related to running the model, such as electricity, server costs, data subscriptions, and trade fees, but before taxes, are subtracted. The earnings on the first day indicated above are after all these expenses were deducted.

I will post weekly updates of the model's progress towards its goal of making the market resolve YES. Are its goals realistic and can it achieve them?

UPDATED: The model's "production entry" date has been set to January 29. The deadline for the bot to achieve its goals is January 28, 2025.

Because of how long it takes to move money, we expect that it will be at least three weeks before all the money is able to be moved to the accounts through the slow ACH system. We are also seeking investment and hope to borrow as much money as possible.

Therefore, rather than track absolute gains, I will post daily percentage gains or losses on however much was bought that day and compound them.

The current adjusted profit (compounded regardless of amount traded) is 17.062% over 45 days the model was active (ending May 31, 2024.) The model was inactive for 65 additional days because a failed motherboard led us to erroneously believe we were receiving poor validation statistics.