Skip to content

Latest commit

 

History

History

TimeSeriesForecasting

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
.DS_Store
.DS_Store
 
 
README.md
README.md
 
 
lstm_stock_model.h5
lstm_stock_model.h5
 
 
tesla_stock.csv
tesla_stock.csv
 
 
time_series_forecast.ipynb
time_series_forecast.ipynb
 
 

README.md

Tesla Stock Market Analysis

Exploratory analysis and trend forecasting of Tesla (TSLA) stock using 5 years of historical price data. The project applies technical indicators and a baseline regression model to understand price behaviour and trend direction.

What This Project Does

Step Technique Purpose
Data collection yfinance API Pull OHLCV data, Jan 2020 – Dec 2024
EDA Summary stats + price chart Understand price distribution and range
Technical indicators 50-day SMA, 20-day EMA Smooth noise, identify momentum
Volatility analysis Bollinger Bands (20-day, ±2σ) Flag overbought/oversold conditions
Trend forecasting Linear Regression (date → close) Establish a naive trend baseline

Dataset

  • Source: Yahoo Finance via yfinance
  • Ticker: TSLA (Tesla Inc.)
  • Period: January 2, 2020 – December 30, 2024
  • Trading days: 1,257
  • Features used: Date, Open, High, Low, Close, Volume
Statistic Close Price
Mean $213.28
Std Dev $83.32
Min $24.08 (Jan 2020)
Max ~$479 (Nov 2021 peak)
25th pct $173.86
75th pct ~$266

Technical Indicators

Moving Averages

  • 50-day SMA — identifies medium-term trend direction; lags price but filters short-term noise
  • 20-day EMA — reacts faster to price changes; useful for spotting momentum shifts earlier

Bollinger Bands

Constructed around a 20-day rolling mean ± 2 standard deviations:

  • Upper band — price approaching here signals potential overbought territory
  • Lower band — price approaching here signals potential oversold territory
  • Width — expanding bands indicate higher volatility; contracting bands suggest consolidation

TSLA is known for wide, volatile Bollinger Bands — periods like late 2020 and late 2021 showed extreme band expansion.

Baseline Model: Linear Regression

A Linear Regression was fit using the trading day index as the sole feature — a naive trend line that captures the overall direction of price across the full 5-year period.

  • Train/test split: 80% / 20% (1,005 / 252 samples)
  • Input feature: Numerical date index (factorized)
  • Target: Closing price

Note: This is intentionally a baseline. A date-indexed regression cannot capture TSLA's volatility, regime changes, or non-linear patterns — but it establishes a trend floor useful for comparison with more sophisticated models.

Key Observations

  1. Price range is extreme — TSLA went from ~$28 in Jan 2020 to a ~$479 peak in Nov 2021, then corrected sharply to ~$110 in late 2022, before recovering above $400 by end of 2024. Linear trend models struggle significantly with this behaviour.

  2. Technical indicators reveal regime shifts — The 50-day SMA and 20-day EMA crossed meaningfully in mid-2021 and again in mid-2022, aligning with major price inflection points.

  3. Bollinger Bands confirm volatility clustering — Tesla has unusually wide bands compared to stable stocks, making mean-reversion strategies risky without additional confirmation signals.

  4. Linear regression captures trend, not volatility — The baseline model fits the general upward slope but misses every peak and trough. This confirms the need for time-series-aware models for actual forecasting.

Tech Stack

Tool Use
yfinance Data download
pandas Data wrangling, rolling calculations
numpy Numerical operations
matplotlib All visualizations
scikit-learn Linear Regression, train/test split

File Structure

TimeSeriesForecasting/
│
├── Tesla_Stock_Market_Analysis.ipynb   # Main notebook
├── tesla_stock.csv                     # Raw price data
├── lstm_stock_model.h5                 # Saved model (future extension)
└── README.md

How to Run

pip install yfinance pandas numpy matplotlib scikit-learn

# Open the notebook
jupyter notebook Tesla_Stock_Market_Analysis.ipynb

Cells run top-to-bottom. Data is fetched live from Yahoo Finance — results may differ slightly if re-run on a later date.

What I Learned

Technical indicators are interpretation tools, not signals. SMAs and Bollinger Bands describe what has happened — combining them meaningfully requires domain context. For a stock like Tesla, which is driven heavily by news cycles and sentiment, technical indicators alone are insufficient for prediction.

Baseline models matter. Building the linear regression first made it immediately obvious why more complex approaches are needed. It's easy to jump to LSTM and claim good results; it's harder to show why a simple model fails and what that reveals about the data's structure.

Random splitting is wrong for time series. The train/test split in this notebook uses sklearn's default random shuffle, which leaks future data into training. A proper time-series split should use a cutoff date (e.g., train on 2020–2023, test on 2024). This is a known limitation of the current implementation.

Next Steps

  • Fix train/test split to use a chronological cutoff (no data leakage)
  • Add proper evaluation metrics: MSE, MAE, RMSE on the test set
  • Implement ARIMA with stationarity testing (ADF test) and parameter selection via ACF/PACF
  • Build LSTM on 60-day lookback windows (framework already scaffolded with lstm_stock_model.h5)
  • Add a confusion matrix / direction-accuracy metric (did the model predict up/down correctly?)
  • Compare all models in a single results table