Forecasting CS2 Skin Volatility
Do textbook volatility models work in a market with no market makers, no arbitrageurs and no short selling?
22/24
items show volatility clustering
0/19
items where GARCH wins
19/19
EWMA in the 90% MCS
The question
CS2 skins trade on the Steam Community Market with none of the plumbing of real markets. That makes it a clean test of whether standard volatility models actually earn their complexity.
What I did
Built a daily price panel of 24 skins across four price tiers. Tested for volatility clustering, then forecast volatility with GARCH, EGARCH and GJR-GARCH against simple benchmarks (EWMA and a 30-day rolling standard deviation), using rolling one-step-ahead forecasts with weekly refits. Compared them with QLIKE loss, Diebold–Mariano tests and the Model Confidence Set.
What I found
Volatility clustering is everywhere (22 of 24 items), yet no GARCH model beat a 30-day rolling standard deviation on any item. EWMA sat inside the 90% Model Confidence Set on all 19 testable items. Bonus: returns follow a weekly cycle, weakest on Wednesdays and strongest on Saturdays.
Next project
Malaysia 2018: A Synthetic Control Study