Using several modeling methods to predict Netflix's stock prices over a set number of years. Each model is iteratively better than the last, culminating with ARIMA.
Authors: Tamara Ponjavic, Narayani Tulsian
Course: MCS 548
This project explores whether stock prices can be predicted with reasonable accuracy by comparing several statistical and time-series forecasting methods on historical Netflix (NFLX) stock data. The broader goal was not just to chase predictive accuracy, but to examine how different models behave when applied to a real financial time series shaped by trend, volatility, and randomness. The project uses Netflix as a case study because its stock history spans multiple business eras — from the early DVD-by-mail years through major growth phases and into the COVID-era market environment.
Can we predict stock prices using model selection with reasonable accuracy? A second question driving the project was whether increasingly sophisticated models perform meaningfully better than simpler ones when faced with the realities of financial data.
The dataset consists of historical Netflix stock prices from 2002 to 2022, sourced from Yahoo Finance. The original data is multivariate, and parts of the analysis focus specifically on variables such as open prices and close prices. Some later experiments also use a narrower time window from 2021-01-02 to 2022-01-02 for short-horizon forecasting.
This project was designed to demonstrate two things:
- how model selection matters in time-series prediction, and
- why even well-chosen models struggle when the underlying process is noisy, non-stationary, and influenced by real-world shocks.
In other words: this was as much a lesson in the limits of prediction as in the mechanics of forecasting.
The project evaluates several models, moving from simpler approaches to more specialized time-series methods:
- Multivariate Linear Regression
- Autoregression (AR)
- Single Exponential Smoothing
- Double Exponential Smoothing
- ARIMA
The presentation also notes that some additional methods were considered and discarded, including Autoregressive Moving Average, some other supervised learning methods, Triple Exponential Smoothing, models relying on trading volume, and deeper-learning-style approaches such as LSTM and XGBoost.
A major part of the project involved understanding the structure of the Netflix stock series before modeling.
Key steps included:
- plotting stock prices across the full 2002–2022 period,
- checking for trend, autocorrelation, and seasonality,
- testing for stationarity,
- differencing the series to reduce non-stationarity,
- reducing the multivariate dataset to a univariate time series for certain models,
- and using a persistence baseline for comparison.
The presentation reports a p-value of 0.613, leading to the conclusion that the original series was non-stationary. Because of this, differencing was used to make the data more suitable for time-series regression methods. The analysis also found strong autocorrelation, a visible trend, and no strong seasonality.
Each method was chosen for a reason:
- Linear Regression was used as a basic benchmark, but it is poorly matched to stock-price time series because it does not naturally handle temporal dependence.
- Autoregression leveraged past values to predict future values, which makes sense when autocorrelation is present.
- Exponential Smoothing was included to test whether weighted recent observations could capture short-term behavior more effectively.
- ARIMA was used as the most flexible of the core models because it can incorporate autoregression, differencing, and moving-average error structure.
The project finds that the models do not perform equally well.
- Linear Regression was effectively a poor fit for the task and was described as not useful in this context.
- Autoregression produced a reported test MSE of 12.558, but its performance was limited by the fact that the stock series was not stationary.
- ARIMA performed best among the tested models and was reported to predict short-term stock prices with relatively strong accuracy. The presentation references ARIMA variants including (52,1,0) and (52,1,1), with reported RMSE values around 11.8–12.0.
- The moving-average component was noted as improving the outcome.
The key conclusion is blunt and correct: stock prices are not very predictable with high reliability, especially over longer horizons.
The project concludes that:
- short-term prediction may be possible,
- long-range forecasting is much less reliable,
- and real financial markets are influenced by too many external and stochastic factors for clean prediction.
This reflects the practical reality of markets more than a modeling failure. A model can capture some structure, but it cannot fully tame randomness.
The presentation frames the difficulty of prediction in realistic terms. Stock prices are affected by:
- market volatility,
- political climate,
- interest rates,
- scandals,
- government actions,
- leadership behavior,
- and other unpredictable external shocks.
The project also references ideas such as the Random Walk Theory and the Fundamental Theorem of Asset Pricing to reinforce the point that market prices are not easily forecastable in a stable, deterministic way.
This project has several important limitations:
- the data is non-stationary in its original form,
- standard regression assumptions break down in time-series settings,
- autoregression alone does not model prediction errors well,
- and strong apparent historical patterns do not guarantee stable future predictability.
Additionally, this repository does not currently include the original notebook used for the analysis. This README is therefore a reconstruction based on the final presentation and is intended to document the project’s goals, methods, and conclusions as faithfully as possible.
This project is useful not only as a stock-price forecasting exercise, but also as a case study in:
- time-series preprocessing,
- model selection under real-world data constraints,
- evaluating forecasting assumptions,
- and understanding the gap between model performance on paper and uncertainty in live markets.