DeepSeek-V3.2发布：性价比再度拉升，金融任务评测表现亮眼

• DeepSeek-V3.2 introduces DSA mechanism for sparse attention DeepSeek-V3.2 incorporates DeepSeek Sparse Attention (DSA) mechanism, which reduces attention complexity from traditional full attention $O(n^2)$ to $O(nk)$, where $k$ represents the number of key tokens selected by the Indexer. This significantly improves efficiency for long-text processing, reducing computational costs while maintaining output quality[13][16][19] - DeepSeek-V3.2 enhances RL framework for tool-use thinking The model optimizes reinforcement learning algorithms, specifically Group Relative Policy Optimization (GRPO), and integrates multi-task training (reasoning, tool-use, human preference alignment) into a single RL phase. This improves performance across complex tasks like mathematics and programming, achieving results comparable to GPT-5 and Gemini-3.0-Pro[22][23][25] - DeepSeek-V3.2 excels in CFLUE financial task evaluations Using CFLUE benchmarks, DeepSeek-V3.2-reasoner and Speciale versions outperform competitors in financial knowledge assessments and application tasks, including ESG classification, financial event extraction, reading comprehension, and text generation. The Speciale version achieves the highest scores in key scenarios[26][27][30] - National 2000 Index enhancement strategy with composite factors The strategy combines factors such as technical, reversal, and residual volatility, neutralized by industry and market capitalization. The enhanced factor achieves an IC mean of 12.63% and a t-statistic of 12.79, demonstrating strong predictive performance[33][34][35] - National 2000 Index enhancement strategy performance The strategy, based on enhanced factors, delivers an annualized excess return of 13.39% and an IR of 1.74. November's excess return was 1.84%, showing recovery in performance[37][40][41] - TSGRU+LGBM machine learning model for index enhancement The model integrates TimeMixer's multi-scale mixing and seasonal/trend decomposition mechanisms into GRU, combined with LightGBM and traditional factors. It captures recent market trends effectively, achieving strong results across multiple indices[43][44][48] - TSGRU+LGBM performance on CSI 300 Index The strategy achieves an annualized excess return of 6.12% and an IR of 1.31. November's excess return was 3.49%, indicating robust recovery[45][48][49] - TSGRU+LGBM performance on CSI 500 Index The strategy delivers an annualized excess return of 9.87% and an IR of 2.01. November's excess return was 1.60%, showing improvement[47][52][53] - TSGRU+LGBM performance on CSI 1000 Index The strategy achieves an annualized excess return of 13.35% and an IR of 2.48. November's excess return was 0.99%, reflecting stable performance[50][54][55] - Dividend style timing and stock selection for fixed-income+ strategy Using 10 macroeconomic and liquidity indicators, the strategy constructs a dynamic event factor system for timing and AI-based stock selection within the CSI Dividend Index. The stock selection strategy achieves an annualized return of 18.81% and a Sharpe ratio of 0.90, outperforming the index[56][57][59]