Sampling-Based Compilation and Forecasting of Implied Volatility Surfaces
ORIGINAL / Compiling and forecasting implied volatility surfaces via survey sampling, tempered inverse-probability sequence learning, and causal activity clocks
Using an NVDA equity and options panel, the study compares stratified and liquidity-weighted PPS sampling for implied volatility surface reconstruction, tests PPS, Neyman and simple random sampling for LSTM training efficiency, and evaluates predictability of prices and forward IV along with activity-clock effects on return distributions.
01 ABSTRACT
The work combines finite-population survey sampling, tempered inverse-probability sequence learning and causal activity clocks to reconstruct and forecast implied volatility surfaces on a split-adjusted NVDA panel. Sampling design matching the target functional drives cross-sectional accuracy, while short-horizon prices do not beat a random walk with drift and forward IV mean-reversion appears only in a few post-shock sessions and is insignificant under clustered tests. Activity clocks are presented as descriptive stabilizers of sampling cadence rather than confirmed improvements.
02 KEY FINDINGS
- Plug-in stratified allocation using lagged cell dispersion improves equal-weighted surface reconstruction with full grid coverage, whereas liquidity-weighted PPS inflates error when misapplied to that target.
- PPS and Neyman designs retain more tail-volatility windows in LSTM training and are equivalent to full-census training within chosen margins while cutting training time, but incremental forecasting gains over simple random sampling are not established.
- Short-horizon prices do not beat a random walk with drift; forward IV mean-reversion is concentrated in a few post-shock reversal sessions, insignificant under session-clustered tests, and sequence models do not outperform a linear HAR-IV benchmark.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV Q-FIN