chat-spx 0.2.0

Close prices memorized in a transformer checkpoint

chat-spx

This small AI package is a chat engine. You send it a date, it tells you the SPX Close on that date.

This package is based on a PyTorch transformer model and is predict-only. The bundled data contains the model weights for frozen trading dates from 5/26/1896 to 5/29/2026. Before 12/30/1927, the closing price is rescaled from the Industrial index.

Internally, the transformer memorizes three prefix tasks: M -> max(date_id), D -> YYYYMMDD, which rebuilds the date_id -> YYYY-MM-DD index at load time, and P -> log_px_fixed, which reconstructs log_px and then px. This structure requires only minimal metadata outside of the model.

Disclaimer: There is no guarantee of data precision since the AI model is prone to error. Use at your own risk.

Install

From this repository:

pip install .

For editable development:

pip install -e .

Usage

from chat_spx import get_px, get_log_px, predict

px = get_px("1927-12-30")
log_px = get_log_px("1927-12-30")
record = predict("1990-01-05")

print(f"{px:.6f}")  # 17.660000
print(log_px)
print(record["pred_log_px_fixed"])

Date ranges return a list of floats for trading dates in the inclusive range:

from chat_spx import get_px_by_date_range

pxs = get_px_by_date_range("2026-05-18", "2026-05-21")
print([round(px, 2) for px in pxs])  # [7403.05, 7353.61, 7432.97, 7445.72]

For repeated calls, create one ChatPX instance so the model is loaded once:

from chat_spx import ChatPX

chatpx = ChatPX()

print(f"{chatpx.get_px('1927-12-30'):.6f}")  # 17.660000
print(f"{chatpx.get_px('1990-01-05'):.6f}")  # 352.200000
print([round(px, 2) for px in chatpx.get_px_by_date_range("2026-05-18", "2026-05-21")])

You can also pass date parts:

from chat_spx import get_px

px = get_px(1990, 1, 5)
print(f"{px:.6f}")  # 352.200000

The command-line entry point prints one price by default:

chat-spx 1990-01-05
chat-spx --price 1990-01-05

For a trading-date range:

chat-spx --range 2026-05-18 2026-05-21

Notes

• get_px(date) returns exp(log_px).
• get_px_by_date_range(start_date, end_date) returns prices for frozen trading dates in the inclusive range.
• get_log_px(date) returns the memorized fixed-point log_px value.
• predict(date) returns the date id, predicted digits, fixed integer, log_px, and px.
• The checkpoint uses prefix queries: P -> log_px_fixed, D -> YYYYMMDD, and M -> max(date_id).
• The date must be in the frozen trading-date index. Weekends, holidays, and dates after 5/29/2026 raise KeyError.

Tests

pip install -e ".[test]"
pytest

Publishing

Build with:

python -m build

Upload with:

python -m twine upload dist/*

The package includes the model weights, so the distribution is much larger than a typical pure-Python package. It is about 21MB.