A comparative forecasting study in Kashgar Prefecture finds that the TimeMixer neural network reduced tuberculosis prediction errors at a nine-month horizon across twelve counties. Naive persistence performed best at shorter three- and six-month windows, suggesting that model selection may need to match the forecasting horizon. The evaluation is notable because operational TB planning depends on horizon-specific reliability, from program resource allocation to intervention timing. Using a consistent multi-county framework also helps distinguish model performance from local quirks. For digital health and implementation teams, the key point is the practical performance boundary of forecasting tools: improve longer-horizon predictions without assuming gains across all timeframes. If adopted, these models could complement surveillance systems by providing more stable projections for longer planning cycles, while retaining simpler approaches for short-term forecasting needs.