Random row splits in an asset panel

Why can a random split of stock-date rows overstate performance when the intended use is forecasting future dates?

Training and test rows can share dates, market shocks, overlapping labels and closely related issuers. The split then measures interpolation among contemporaneous observations rather than genuine future forecasting. Split chronologically and keep appropriate groups together. If generalisation to unseen assets is also required, add an asset-held-out evaluation. The right split follows the deployment question; one split does not answer every generalisation claim.