duemig /Stanford-Project-Predicting-stock-prices-using-a-LSTM-Network

CS230 (Stanford) project: LSTM networks vs. regularised logistic regression for the monthly direction of the S&P 500, judged against the base rate

★ 299⑂ 50👁 12⑃ 0● 3▤ 0
Jupyter NotebookCreated 2019-02-19
+0
日增长
299
stars
50
forks
12
watchers
0
prs
3
issues
0
releases

README

Predicting stock prices with LSTM Networks

Can a long short term memory (LSTM) network predict whether the S&P 500 will be up
or down over the next month, and does it do any better than a regularised logistic
regression or a gradient-boosting model? The short answer is no. Across every
feature set the models score close to the naive rule of always predicting "up",
and a timing strategy built on their forecasts does not beat buying and holding
after costs.

It started as my course project for CS230 Deep Learning at Stanford.

The written report is report.pdf and the notebook is
Final Project.ipynb. Every number below is read from the
files in results/, which are produced by run_study.py.

 Method

- Target. One non-overlapping observation per month, labelled by the sign of
  the next month's return. This removes the heavy overlap of the original daily
  labelling.
- Features. Stationary features over 23 look-back windows for each of three
  families: summed log returns, rolling volatility, and log volume relative to its
  trailing one-year mean (which removes the long-term growth in turnover). The
  three families give the seven feature sets of the original project. The baselines
  use the 69 month-end aggregates; the LSTM reads the most recent 60 trading days
  of the daily channels.
- Validation. The last 20% of months are held out as a test set, used once.
  The rest is tuned with purged walk-forward cross-validation: a training month
  whose label reaches a validation month is purged, and a one-month embargo is
  dropped before each validation block. The baselines fit their scaler inside each
  fold; the LSTM standardizes its inputs on the training part of its inner split,
  so no scaling ever sees the data it is evaluated on.
- Models. Always-up (the benchmark), an elastic-net logistic regression and a
  gradient-boosting classifier on the aggregates, and an LSTM on the daily
  sequences trained with early stopping over 10 seeds.
- Evaluation. Accuracy against the base rate, AUC and Brier score with 95%
  moving-block bootstrap intervals, a paired comparison against always-up, and a
  timing strategy (long the index next month when the predicted up-probability
  exceeds a threshold chosen on validation data only) against buy-and-hold after
  costs.

 Data

The data are daily S&P 500 quotes (^GSPC) from 1980 to the end of 2025, about
11,595 trading days. The price data is not committed. run_study.py downloads
it with yfinance on first use and caches it under data/; the source, coverage
and download date are recorded in results/summary.json. If
Yahoo is blocked, the code falls back to Stooq. To use a file you downloaded
yourself, save it as data/gspc.csv with a Date index and Close and Volume
columns.

 Results

The test set is 106 months (February 2017 to November 2025) and is 68.9% up
months, so always predicting "up" already scores 68.9%. Test-set accuracy and AUC:

| Set | Features | Acc EN | Acc GB | Acc LSTM | AUC EN | AUC GB | AUC LSTM 

Top issues

Snapshot: 2026-11-16T09:23:14Z · Data license Apache-2.0 · JSON