A watchlist records a form of attention prior work could not see: a self-authored, dated commitment to follow a stock, made before any purchase. Using 28,990 Chinese retail investors whose every purchase is linked to a timestamped watchlist entry, I show that the stocks investors watch but do not buy outperform the stocks they buy — yet among purchases, those routed through the watchlist (warm) underperform purchases made on sight (cold) by 234 basis points of characteristic-adjusted return over the next sixty trading days. The list’s signal is good; the trap is in the conversion from watching to buying. Roughly 45% of the gap reflects composition — watched purchases carry more costly speculative tendencies. The surviving mechanism for the rest is reference-dependent conversion: the daily first-purchase hazard changes slope sharply as the price crosses the level at which the stock was added and the extremes witnessed while it waited.

Why does the disposition effect survive in information-rich trading environments? Using clickstream and trading records for 10,000 investors on a large Chinese online trading platform, we find that active information collection sharpens gain-side exits but worsens loss-side outcomes: clicked-on retained losers underperform by an additional 1.5–2.0 percent, offsetting the net trading benefit. Two mechanisms drive the asymmetry: an ostrich effect in acquisition, whereby focal-stock losses reduce abnormal clicking by 39 percent, and cognitive inertia in processing, whereby investors who examine losses fail to act. Because the preferences that generate the bias also allocate attention, self-directed information collection cannot correct it.

On February 8, 2025, a large Chinese retail brokerage swapped the large language model behind its in-app chatbot for DeepSeek, leaving the chatbot’s sibling services untouched. Measuring exposure to the upgrade with strictly predetermined pre-launch usage, I find that more-exposed investors consulted the news less often, while the composition of what they still read — concentration, slant, staleness, pre-trade diligence — is bounded near zero: the upgrade changed how often investors read, not what. Around the same upgrade, the risk-adjusted performance of these investors’ subsequent first purchases deteriorated relative to observably identical engaged users — a within-stock, mistimed-entry effect that survives reversion controls, permutation inference, and nine placebo launch dates. As AI assistants become a front door to financial information, a quality upgrade can change when investors act without changing what they know.

We study how asset ownership affects consumer behavior using fractional music-royalty auctions on a Chinese online platform where randomly selected users bid for song royalties. In a difference-in-differences design comparing auction winners to non-winning bidders, obtaining ownership of a music asset significantly raises the owner’s subsequent play count of the auctioned song and overall music consumption, especially for songs in the same genre. Financial incentives cannot explain these effects: royalty income earned from the asset bears almost no relationship to the change in consumption. Consistent with behavioral theories in which ownership elevates an asset’s subjective value and attracts attention, the effect is strongest for engaged platform users, publicly visible consumers, and assets aligned with pre-existing genre preferences — field evidence of an endowment effect in digital music consumption, inducing demand beyond what economic incentives predict.

Generative AI increasingly assists the creation of cultural goods, and regulators are mandating that AI involvement be labeled — on the premise that a label taxes the artists who use it. Does disclosing AI assistance actually cost artists, or does the apparent penalty reflect the strategic choice of which work they reveal? We study a leading music-streaming platform whose integrated AI tool let artists choose whether to credit AI use, before a platform-wide mandate removed that choice. Under voluntary disclosure, concealed AI songs outperform the same artist’s credited AI songs roughly threefold in plays — but artists credit AI on songs that enter new genres or carry a more visible human hand. Exploiting the mandatory-label rollout in a difference-in-differences design, we estimate the causal listener penalty at about five percent of plays: observational data overstate the market cost of mandated AI disclosure by more than an order of magnitude.

Attention is a double-edged sword in financial markets: it motivates investors to research stocks and trade on what they learn, potentially improving price discovery, yet it can also discourage research and trading when investors believe the news is already priced in or when strategic substitutability makes them hesitant to act. I examine this dual role through mobile push notifications, whose delivery is otherwise endogenous: recommender systems push news about the very stocks investors already track. Exploiting a random technical split under which a small share of a brokerage’s push notifications goes undelivered, I compare investors who received a push — the news plus an attention trigger — with investors who received the news only. A push raises active stock research by 2.7 percent within the hour, far more for watchlist, portfolio, and focus-list stocks, and the effect dissipates within a day.