George Liu

Markets & analysis · Active

Prediction Markets ↔ Financial Markets

Exploring information flow between prediction markets and traditional financial instruments — signals, lags, and structural differences.

Executive summary

Prediction markets and financial markets both aggregate beliefs under uncertainty, but they differ in participants, incentives, liquidity, and contract design. This project investigates how information appears to move between them — as an engineering and statistical problem, not a promise of alpha.

TODO: Link only to repositories and write-ups you actually maintain. Remove any URL that is not yours. Do not invent backtest returns.

Problem / opportunity

Event contracts (elections, macro prints, corporate outcomes) often reprice when traditional markets are closed, fragmented, or slow to express a narrative. Conversely, liquid financial instruments can lead niche event markets. Understanding those relationships requires careful data plumbing and humility about causality.

Context

Independent research interest spanning statistics, systems, and market microstructure. Adjacent work includes market-memory style context stores and automated monitoring bots — tooling for observation, not managed products.

Role

Design of data collection, storage, and analysis workflows; writing that separates signal hypotheses from confirmed effects.

Constraints

  • API rate limits and historical completeness vary by venue
  • Event definitions rarely map 1:1 onto tickers
  • Survivorship and resolution rules distort naive comparisons
  • No confidential or non-public data sources

Approach

  1. Define a small universe of events with clear resolution criteria
  2. Ingest prediction-market prices and relevant financial series on a common clock
  3. Store raw and normalized readings with provenance
  4. Compute simple transforms: returns, probability changes, lead/lag windows
  5. Document hypotheses before peeking at results
# Illustrative sketch — not a live trading system
def align_series(event_prob, asset_return, freq="1h"):
    """Resample and join on UTC timestamps; forward-fill only within policy."""
    ...

Difficult problems

  • Semantics: “Will X happen?” is not the same as “ticker Y moves”
  • Latency & staleness: Comparing last trades across venues without microstructure context misleads
  • Multiple testing: Easy to find a story in noise if you hunt long enough

Tradeoffs

Building a robust research lab (storage, lineage, notebooks) competes with shipping flashy dashboards. This project prefers correct timestamps and boring reliability over UI polish.

Outcome

TODO: Summarize findings only after you have written them up honestly. Placeholder: Work in progress — tooling first, claims later.

What I learned

  • Information-flow questions force you to get serialization, time zones, and missing data right
  • Markets are social systems; pure price series without context is incomplete
  • Automation helps consistency more than cleverness helps insight

What I would change

  • Stronger event ontology from day one
  • Explicit “hypothesis registry” to reduce storytelling bias
  • Clearer separation between research code and any future product surface

See notes under tags prediction-markets and information-flow on the Writing page.