01RanTime series

Apple price with an LSTM

I downloaded about 3,400 trading days of Apple prices and trained a two-layer LSTM to output the closing price from the same day's open, high, low and volume. The rows were shuffled before the 80/20 split.

Original tutorial by Aman Kharwal: Stock Price Prediction with LSTM Neural Network (opens in a new tab)

What it printed in 2022

Numbers read from the saved notebook outputs. Cell numbers count every cell in the notebook, Markdown included, starting at 1.
Output	Value	Cell #
Training MSE after epoch 30	3.62	21
LSTM parameters	117,619	20
Prediction for one hand-typed day (USD)	177.07	23

Cell # counts every cell from the top of the notebook, Markdown included, starting at 1.

The 3.62 is the training loss after the last epoch. The held-out rows were never scored, so the notebook says nothing about how well the model predicts.

The data

Daily AAPL prices from Yahoo Finance through the yfinance package, 2009 to 2022.

What I'd do differently

Predict tomorrow from information available today. Feeding the same day's high and low into a model of the same day's close leaks most of the answer.
Split by date, train on the past and test on the future, and never shuffle a time series.
Compare with 'tomorrow's close equals today's close' before calling any model good.
Scale the inputs. Volume sat in the hundreds of millions next to prices below 200.

Re-checked in the audits

01 · 05 · 09 · 28 · 32
Conclusion flips
Stock forecasts against 'tomorrow equals today'
On a time-ordered split the Apple LSTM misses the close by 38.7% on average, against 1.7% for 'tomorrow equals today'. The linear model in notebook 05 also loses to 'no change' (0.50% versus 0.36%).
Read the audit

Apple price with an LSTM

What it printed in 2022

Numbers read from the saved notebook outputs. Cell numbers count every cell in the notebook, Markdown included, starting at 1.
Output	Value	Cell #
Training MSE after epoch 30	3.62	21
LSTM parameters	117,619	20
Prediction for one hand-typed day (USD)	177.07	23

Cell # counts every cell from the top of the notebook, Markdown included, starting at 1.

The 3.62 is the training loss after the last epoch. The held-out rows were never scored, so the notebook says nothing about how well the model predicts.

What I'd do differently

Predict tomorrow from information available today. Feeding the same day's high and low into a model of the same day's close leaks most of the answer.

Split by date, train on the past and test on the future, and never shuffle a time series.

Compare with 'tomorrow's close equals today's close' before calling any model good.

Scale the inputs. Volume sat in the hundreds of millions next to prices below 200.

Re-checked in the audits

01 · 05 · 09 · 28 · 32

Conclusion flips

Stock forecasts against 'tomorrow equals today'

On a time-ordered split the Apple LSTM misses the close by 38.7% on average, against 1.7% for 'tomorrow equals today'. The linear model in notebook 05 also loses to 'no change' (0.50% versus 0.36%).

Read the audit