50 time series dissertation questions grouped by modelling approach, from seasonal decomposition and ARIMA to cointegration, VAR, GARCH, regime switching and neural forecasting.
By Siddharth Gupta · Statssy
Time series analysis supports work in economics, finance, public health, environmental science and operations. The method should be selected for the question and the available observation frequency—not simply because it is more complex.
These 50 ideas pair each dissertation question with a model, data structure, public source and related thesis. Pay particular attention to chronological validation, nonstationarity, structural breaks and whether the linked source actually contains the required observations.
Match the structure of the data and the question to the model family, then check assumptions and sample length.
| Your research question | Starting approach |
|---|---|
| Is there a long-term trend or seasonal pattern? | Trend, seasonality and decomposition |
| What will the next values be? | ARIMA and SARIMA forecasting |
| Which forecasting method performs best on unseen observations? | Forecast comparison and evaluation |
| Do two economic series share a long-run relationship? | Unit roots, cointegration and error correction |
| Does one time series help predict another? | VAR, Granger causality and impulse responses |
| Does uncertainty or volatility change over time? | Volatility models (ARCH and GARCH) |
| Did a policy or event change the level or trend? | Interrupted time series and event analysis |
| Has the series changed its behaviour across regimes? | Structural breaks and regime switching |
| Can machine learning outperform statistical forecasting? | Machine learning and deep learning forecasting |
| Do many interconnected series have useful predictive dependencies? | Multivariate, causal and applied time series |
Start by describing the series. These topics test if a trend is real and how strong the seasonal pattern is.
Data check: This COVID-19 source tracks a specific infection, not all respiratory illnesses; define COVID-19 as the outcome or obtain disease-specific surveillance data.
Data check: NOAA Climate Data Online mainly provides land-based weather observations. Coastal upwelling needs dedicated sea-surface-temperature, wind-stress or oceanographic records.
Data check: Water Quality Portal is typically sampled intermittently; daily cycles require a sensor site with sufficiently frequent continuous nitrate measurements.
ARIMA is the most common forecasting model in dissertations. These topics use it on real monthly and daily series.
Data check: Check VIX series and VIX derivatives’ launch dates. ARIMA predicts the specified index or returns, not latent realised volatility by itself.
Data check: Use a BTS freight series that actually records monthly cargo; the linked ONTIME page focuses on airline punctuality.
A forecast only matters if it beats the alternatives. These topics compare methods and test accuracy properly.
Data check: WASDE contains USDA forecasts, not individual private analysts’ forecast histories. A separate private-consensus series is required.
Data check: Neither WASDE nor CFTC gives complete private analyst forecasts or tradable daily settlement prices; both are needed for a trading simulation.
Data check: Kenneth French portfolio returns are not individual stock-level betas; specify a portfolio-beta study or add stock data.
When two series share a long-run path, cointegration and error correction models show how they adjust.
Data check: Confirm land lease rates and matching farmland-value series are available for the same geography and years.
Data check: USDA WASDE does not directly report observed Southeast Asian rice export transaction prices. Obtain comparable export and FX prices.
Vector autoregressions model several series together. They show which series lead and how shocks spread.
Data check: Monthly and quarterly economic series must be aligned before fitting a VAR; temporal aggregation can change lag interpretation.
Data check: Match Bitcoin daily returns to monthly macroeconomic indicators without leaking lower-frequency information into past dates.
Data check: CFTC position/open-interest records may not separate rice contracts at the requested granularity or include the full crop calendar.
Financial series have calm and wild periods. GARCH models capture that changing volatility.
When a policy or event happens on a known date, interrupted time series shows the change in level and trend.
Data check: A controlled interrupted time series needs a comparison location or pollutant and an independent lockdown timing series; EPA AirData alone is not sufficient.
Data check: Daily cotton futures prices are not included in WASDE or CFTC position totals. Add settlement or intraday prices.
Data check: Google Trends measures search interest, not diagnosed psychological distress; define the outcome accordingly and source hurricane exposure dates.
Data check: USDA WASDE and weekly export-sales releases are distinct reports. CFTC positions also are not commodity settlement prices; obtain the correct release schedule and returns.
Some series change behaviour after a crisis or a new policy. These models find when and how.
Data check: Futures basis needs cash/spot and futures prices for the same commodity and location; positions and WASDE forecasts are insufficient.
Deep learning can beat classic models on long, complex series. Always compare it with a simple benchmark.
Data check: Extreme rainfall is a rare event; compare event-sensitive metrics as well as continuous forecasts and use strictly chronological validation.
Data check: Graph neural networks need a justified graph over customers; the electricity dataset supplies loads, not a physical connectivity graph.
Data check: Google Trends search volume is a proxy for attention, not text sentiment. Use an additional sentiment corpus or change the feature label.
Data check: The linked ONTIME table is not the BTS Origin and Destination Survey fares table; obtain DB1B airfare data for route-quarter averages.
These are advanced topics. They suit students who want to push beyond standard forecasting.
Data check: PCMCI can identify conditional dependencies under assumptions; it does not by itself establish causal effects from observational weather and pollution data.
Data check: The two weather/air monitoring portals may not label dust-storm events explicitly. Create and validate an event label source.
Data check: Rainfall records alone do not identify landslide occurrence. Merge a verified landslide inventory.
Data check: Prediction interval coverage and statistical tolerance intervals have different interpretations; specify the interval estimand before evaluation.
The following free sources support time series research. Ensure that each source supplies the correct frequency and target variable; a multivariate or causal question may require merging another series or an event calendar.
Macroeconomic and financial time series, with observation frequencies and revision histories varying by series.
Historical NSE index values for Indian financial-market research.
Indian inflation, interest rates, exchange rates and other central-bank time series.
Station-based temperature, precipitation and weather data; oceanographic phenomena need other sources.
US pollution monitor observations including PM2.5, ozone and other pollutants.
Energy prices, generation, production and fuel markets at different frequencies.
Quarter-hourly electric power consumption for hundreds of clients.
Large multi-frequency forecasting benchmark and methods; confirm the correct source files and horizon.
Reported futures trader positions and open interest; not a full price-history database.
Monthly agricultural supply-and-demand forecasts, not proprietary analyst forecasts.
Portfolio returns and financial risk factors, not individual company returns.
Normalised search interest by topic, time and geography; not a direct sentiment score.
US airline on-time and operations records; fare series require other BTS survey tables.
Archived COVID-19 cases, deaths and related indicators, subject to source coverage.
Time-stamped water samples from multiple agencies; sample spacing is often irregular.
US underlying or multiple-cause mortality statistics with suppression and period-specific coverage.
The printable guide will contain all 50 questions, study designs, suggested time series models, dataset links and related university thesis references. The download will be added after the PDF is produced and uploaded.
PDF download coming soonFor adjacent topics, explore panel data dissertation topics, finance research topics and economics research topics.
Bring your dataset, frequency and research question to a free 15-minute call. We can review forecasting horizons, model assumptions and evaluation plans.
Book a free 15-minute call15 minutes · No cost · No pressure
Time series analysis studies data recorded in time order, like monthly sales or daily prices. It finds trends and seasons, and it forecasts what comes next.
Start with ARIMA or exponential smoothing as a benchmark. Use machine learning only if it beats them on out-of-sample data.
Aim for at least 50 points for ARIMA, and at least three full cycles for seasonal models. Deep learning needs thousands.
A stationary series has a stable mean and variance. Most models need it, so test with ADF or KPSS and difference the series if needed.
R with the forecast and fable packages is the easiest. EViews and Stata suit econometric models. Python suits deep learning.
Yes. Download the full Topic and Method list from this page. It includes every topic, model and dataset link.