A preprint finds a learned volatility network had lower forecast errors for 20 UK firms than preset structures in a 500-day test.