Dividend Prediction with LSTM
Forecasting future dividends using a sequence model (LSTM) on historical market data. Emphasis on clean preprocessing and time-aware evaluation.
Overview
Built an end-to-end pipeline that fetches market data, prepares supervised sequences, and trains an LSTM to predict upcoming dividend values. The focus was on reliable preprocessing and honest, time-respecting validation.
Data Cleaning
- Loaded price/dividend data with
yfinance; unified to a single datetime index. - Handled missing values and ensured the series were sorted and aligned by date.
- Constructed the target dividend series and filtered to the relevant ticker/time span.
Exploratory Analysis
- Plotted dividend history and matching price series to understand payout patterns.
- Reviewed basic descriptive statistics and missing-value patterns.
Modeling
- Framed the problem as supervised sequences (look-back window → next dividend value).
- Created training samples by sliding windows over the historical series.
- Split data by time (train → validation/test) with no shuffling.
- Implemented an LSTM in Keras and trained on the prepared sequences.
- Evaluated on a held-out time segment and visualized predictions vs actuals.
What I Focused On
ReliabilityClean alignment and consistent indexing
ReproducibilityDeterministic splits and saved seeds
ReadabilityClear plots comparing predicted vs. true dividends
Screenshots
