Inspiration
Both trading and travel data point to the same people: casino "whales". High rollers fly private, and a large share of Las Vegas Strip profit comes from a small number of VIP baccarat players. We asked a simple question: do private-jet arrivals in Las Vegas tell you something about casino stocks before the casinos report? The idea is to trade Strip operators that depend on VIP play (MGM, Wynn) against local-market casinos that do not (Red Rock, Boyd).
What it does
Reads a point-in-time monthly datasets from FAA data OPSNET and TFMSC jet counts at Las Vegas versus the rest of the US as a control, Nevada Gaming Control Board Revenue releases, LVCVA visitor data and SEC filings.
How we built it
- Python and C++: pandas, NumPy, SciPy, and scikit-learn in Python. A C++17 trainer and inventory kernel match the Python versions bit for bit.
- GPU: a correlation sweep over 31,758 cells with 100,000 permutations each, using CuPy.
- Market data: 1-minute and 1-second Databento bars, plus the full order book for 2025-26 liquidity checks ($74 of a $125 budget).
Overfitting checks: a 12-lens look-ahead audit with 21 fixes, deflated Sharpe, probability of backtest overfitting, Benjamini-Hochberg multiple-testing control, walk-forward selection, and a holdout lock with frozen code hashes. Workflow: many AI agents worked in parallel (builders plus independent verifiers), and we checked every result against the files that produced it.
Challenges we ran into
Flight data. OpenSky and other databases with exact per-aircraft flight data hit credit and rate limits. That left our model on the FAA's monthly counts, which limited how fast the signal can react.
Signal decay. Monthly data means the signal decays quickly. Our medium-frequency version had to rely on FAA releases, and no decay setting survived multiple-testing correction.
Publication timing. Avoiding look-ahead meant tracking the exact date and time every FAA, NGCB, TSA, and airport figure became public.
The holdout. COVID closures had to be handled by a rule known in real time. A take-private bid for MGM landed in the middle of the holdout. Costs. Market impact turned a small edge negative.
Accomplishments that we're proud of
- An honest, fully pre-registered test with a one-time holdout, no tuning on out-of-sample data, and every number traceable to code.
- C++ and Python engines that agree exactly, a GPU permutation sweep, and a full cost model checked against the real order book. A clear answer, even though it is negative.
What we Learned
*Trading costs matter more than the signal. A round trip costs about 57-72 bp, mostly market impact.
- Data timing is everything. Most "alt-data edges" disappear once you use information only on the day it becomes public. Sizing helps but cannot rescue a weak signal. Sizing by a percentage of volume lost much less than a fixed dollar size, but none of the 693 settings was profitable after costs. ## What's next for Evaluating flight data to predict Casino Stock prices Grabbing real time data
Log in or sign up for Devpost to join the conversation.