Skip to content
Armand Notue

Machine Learning · Time Series · Deep Learning

Stock Forecasting: LSTM vs ARIMA — Deep learning against classical baselines

A leakage-free comparison of an LSTM, ARIMA and a moving-average baseline on five years of five large-cap stocks — showing that the “best” model depends on the metric.

Complete
Final dashboard: Tesla 2024 forecasts against actual prices, MAPE and directional accuracy for each model and stock, LSTM accuracy per stock, LSTM error versus volatility, and key findings.From the notebook · Final dashboard
Problem
Forecasting demos often report impressive accuracy from leaky pipelines. The honest question is narrower: does a deep model add anything over classical baselines when evaluation is strictly out-of-sample?
Approach
Built 15 technical indicators, split strictly by time (2020–2023 train, 2024 test), and compared a 7-day moving average, auto-ARIMA and a dual-head PyTorch LSTM on price error and directional accuracy.

Key results

LSTM directional accuracy · vs 45.8% ARIMA
55.5%
ARIMA MAPE — lowest price error
1.50%
LSTM parameters, trained in 3.4 s on a T4
126K

01

Overview

A forecasting study on real market data for Tesla, Apple, Amazon, Google and Microsoft (2020–2024, via the Yahoo Finance API), comparing a deep sequence model against classical baselines under a strictly temporal evaluation.

02

Problem & motivation

Most stock-prediction notebooks leak the future into training and report numbers nobody could trade on. This project does the boring part properly — and then asks which model actually wins, on which metric.

03

My role

  • Independent project — data pipeline, feature engineering, modeling and evaluation.

04

System architecture

  1. Stage 1:Data

    • 5 tickers · 5 years

      Adjusted daily prices and volume via yfinance

  2. Stage 2:Features

    • 15 indicators

      Moving averages, RSI, MACD, volatility, lagged returns

  3. Stage 3:Split

    • Temporal split

      Train 2020–2023 · test 2024, no shuffling

  4. Stage 4:Models

    • MA7 baseline

      7-day moving average

    • ARIMA

      Order chosen by auto-ARIMA

    • LSTM

      2 stacked layers, 126,050 parameters, dual heads

  5. Stage 5:Evaluate

    • MAPE

      Price error

    • Directional accuracy

      Up / down, per stock and on average

One pipeline, three models, two metrics, five stocks.

05

Technical implementation

  1. A zero-leakage pipeline

    Every feature uses only information available at prediction time, and the split is strictly by date — the models never train on the period they are tested on.

  2. A dual-head LSTM

    A 20-day window of 15 features feeds LSTM layers of 128 and 64 units and a 32-unit dense layer, then two heads: one regresses the next-day return, one classifies its direction. Trained with early stopping on a Kaggle T4 GPU in 3.4 seconds.

  3. Baselines that are hard to beat

    A 7-day moving average and auto-ARIMA give the LSTM something real to beat. Auto-ARIMA selected order (0,0,0) on Tesla — the signature of a random walk in daily returns.

06

Engineering decisions & challenges

  • Report both metrics

    MAPE and directional accuracy tell different stories. Publishing only one would have crowned a different winner.

  • Per-stock evaluation

    Averages hide volatility. Tesla's daily volatility (4.2%) is more than twice that of the other stocks, and it is the hardest to forecast.

07

Evaluation & results

Averages across the five stocks on the 2024 test period, from the repository README.

LSTM directional accuracy+5.5 points over a coin flip
55.5%
ARIMA MAPELowest price error
1.50%
LSTM MAPESecond on price error
1.90%

The metric picks the winner

ARIMA has the lowest price error; the LSTM is the only model that beats a coin flip on direction. Averages across five stocks, 2024 test period.

Mean absolute percentage errorLower is better
  • LSTM1.90%
  • ARIMA1.50%
  • MA7 baseline2.62%
Show data table
ItemValue
LSTM1.90%
ARIMA1.50%
MA7 baseline2.62%
Directional accuracyHigher is better
  • LSTM55.5%
  • ARIMA45.8%
  • MA7 baseline48.6%

Reference line: Coin flip at 50.

Show data table
ItemValue
LSTM55.5%
ARIMA45.8%
MA7 baseline48.6%

Source: Repository README results table

  • Educational project — not financial advice.
  • Honest evaluation produces modest but trustworthy numbers: the LSTM's directional edge is real across stocks, but small.

08

From the notebook

Real figures from the project's notebook, as published in the repository. Select one to enlarge it.

09

Lessons learned

  1. The metric determines the winner

    Decide what question the model must answer before comparing models.

  2. Leakage is the biggest risk

    A strict temporal split costs impressive-looking numbers and buys trustworthy ones.

10

Technologies used

  • Python
  • PyTorch
  • statsmodels
  • pandas
  • NumPy
  • yfinance

Want the deeper walkthrough?

I'm happy to go through the code, the trade-offs and what I'd do next.

Get in touch