How to use this list
Decentralised finance (DeFi) refers to lending, trading and other financial services run by smart contracts on public blockchains rather than by traditional intermediaries. Because every transaction is recorded on chain, DeFi offers researchers unusually detailed public data, which has encouraged work that combines machine learning with blockchain analytics. A recent review of machine learning for financial market prediction notes both this promise and the persistent problems of black-box models, overfitting and data quality (Rohan et al., 2026).
The rows below keep the sound ideas and add one on regulation. Each row gives a possible research question, a likely design and why it matters now. For wider topics see finance dissertation topics; for the ethics of automated trading see ethical AI in trading systems.

Theme 1: Forecasting with on-chain data
| Topic | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| Do on-chain features improve DeFi forecasts? | Do on-chain measures (for example flows into lending pools or active addresses) improve out-of-sample forecasts of token returns or volatility beyond market price data alone? | Rolling-window forecasting with machine-learning and hybrid models compared with naive and statistical benchmarks; forecast-accuracy tests | ML forecasting studies often overfit and report weak benchmarks (Rohan et al., 2026; Hyndman & Athanasopoulos, 2021) |
Evidence that the gap is real. Rohan et al. (2026) review machine-learning approaches to stock price forecasting from 2018 to 2024 and highlight interpretability, overfitting and data-quality limits, as well as emerging blockchain integration. Standard forecasting practice requires out-of-sample evaluation on rolling windows and comparison with simple benchmarks (Hyndman & Athanasopoulos, 2021). A DeFi study that applies these rules honestly, and reports when complex models fail to beat simple ones, adds more than a new "intelligence framework".
Theme 2: Risk, exploits and market stability
| Topic | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| Early warning of exploits and liquidity stress in DeFi protocols | Can patterns in on-chain activity give early warning of protocol exploits or sudden liquidity withdrawals? | Event dataset of past incidents; classification with time-aware validation; precision-recall evaluation | Incidents are rare, so evaluation must suit imbalanced data (Saito & Rehmsmeier, 2015) |
| Liquidations and volatility in DeFi lending | Do automated liquidations in lending protocols amplify price falls during market stress? | Event study or time-series analysis of liquidation and price data across stress episodes | Lending protocols liquidate positions automatically, so stress episodes can be observed directly in public data |
Evidence that the gap is real. Exploits and crises are rare compared with normal activity, so a model that predicts "no incident" almost always will look accurate. Saito and Rehmsmeier (2015) show that precision-recall analysis is more informative than ROC analysis for such imbalanced problems. For stability questions, clearly defined stress episodes and transparent identification of effects matter more than model complexity. Build the incident list from documented sources, record the date each incident became public, and make sure that no feature used for prediction is measured after that date. Without this discipline an early-warning model can appear to work only because it has, in effect, seen the answer.
Theme 3: Governance and regulation
| Topic | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| Participation and concentration in on-chain governance | How concentrated is voting power in the governance of major DeFi protocols, and is concentration associated with the kinds of proposals that pass? | Descriptive and panel analysis of on-chain voting records; concentration indices | Transparent governance is a central claim of DeFi and can be measured directly |
| DeFi and the EU Markets in Crypto-Assets Regulation | How do MiCA's rules for crypto-asset service providers apply to services with no identifiable intermediary, and what do practitioners expect? | Doctrinal legal analysis plus interviews with compliance staff and developers | MiCA created an EU framework for crypto-assets and their service providers (European Parliament & Council of the European Union, 2023) |
Evidence that the gap is real. Regulation (EU) 2023/1114 established an EU-wide regime for issuers of crypto-assets and crypto-asset service providers (European Parliament & Council of the European Union, 2023); how such rules relate to protocols run by code and token holders is an open legal and policy question. Research on AI tools in conventional finance shows that trust, transparency and perceived fairness shape adoption (Yaseen & Al-Amarneh, 2025), which suggests similar questions for automated tools in DeFi governance.
How to test a topic before you commit
Use the general method in how to choose a PhD topic and how to select a PhD topic, or our topic selection support, and screen each idea with the FINER criteria (Hulley et al., 2013). For DeFi topics, check that you can extract and clean the on-chain data you need, that your sample covers more than one market regime, and that your benchmark models are specified before you see the results.
A short proposal for any of these topics should state the following clearly.
- Scope: which chains, protocols and period, and why they represent the question.
- Target: the exact variable you forecast or detect (returns, volatility, an incident), measured how.
- Benchmarks: the simple and standard models your method must beat (Hyndman & Athanasopoulos, 2021).
- Validation: rolling or time-ordered splits, with no information from the test period used in tuning.
- Metrics: forecast-error measures for forecasting; precision and recall for rare events (Saito & Rehmsmeier, 2015).
- Limits: what pseudonymous data cannot tell you, such as who is behind an address.
Worked example (illustrative). A researcher starts with the live title "AI-driven investment intelligence framework integrating blockchain analytics". She narrows it to one question: do daily net inflows to the largest lending pools improve one-day-ahead volatility forecasts for the protocol's token? She compares a gradient-boosting model, a standard volatility model and a naive benchmark on rolling windows over three years. If the complex model does not win, that is still a publishable finding.
Common mistakes
- Comparing a complex model only with another complex model, never with a simple benchmark.
- Tuning models on the test period, which leaks future information.
- Reporting accuracy for rare events instead of precision and recall.
- Treating on-chain data as clean without handling pseudonymous and automated activity.
- Promising "policy recommendations" without a legal or policy method.
Key takeaways
- Machine learning for market prediction faces known problems of overfitting, opacity and data quality (Rohan et al., 2026).
- Forecasting contributions must beat simple benchmarks out of sample (Hyndman & Athanasopoulos, 2021).
- Rare-event detection should be judged with precision and recall (Saito & Rehmsmeier, 2015).
- Regulation and governance are researchable strands, not afterthoughts (European Parliament & Council of the European Union, 2023; Yaseen & Al-Amarneh, 2025).
- Screen every shortlisted topic for feasibility, novelty, ethics and relevance (Hulley et al., 2013).
