Guosen Securities(SZ:002736) Guoxin Securities·2025-12-14 11:57Core Conclusions - The large models possess certain technical analysis skills, with both DeepSeek V3.2 and Gemini 3 Pro able to perform tasks such as identifying tops and bottoms, drawing segments, and constructing central structures under appropriate prompts [1][2] - For "established" trends, the large models demonstrate a degree of technical analysis capability, with DeepSeek excelling in language organization and long text generation, while Gemini accurately identifies "central expansions" and "trend ambiguities" [1][2] - Gemini has an advantage in "ease of use," as its Nano Banana Pro can perform simple graphic annotations, making it slightly more convenient in practical applications [1] Evaluation Methodology - The evaluation of DeepSeek and Gemini's technical analysis capabilities follows the "Four Consistency Principles," ensuring data source uniformity, identical prompts, concurrent testing environments, and unified assessment standards [2][17] - The models are tested on standardized OHLC price data from the Shanghai Composite Index, with the same task instructions and evaluation criteria based on the original principles of the Chan theory [2][17] Technical Analysis Capabilities - Both models can accurately identify relationships and patterns in K-line data, with Gemini 3 Pro showing a slight edge in recognizing complex structures and providing clear outputs [3][12] - In analyzing established trends, both models demonstrate systematic capabilities, but there are discrepancies in defining fine concepts, particularly in classifying central structures and trend types [3][12] - Gemini 3 Pro is noted for its superior ability to capture the core logic of "divergence + central" in short-term predictions, aligning closely with actual market movements [3] Performance Comparison - The report compares the performance of DeepSeek and Gemini in various aspects of technical analysis, including K-line inclusion processing, top and bottom identification, segment classification, and overall analysis coherence [40] - DeepSeek identified 8 tops and 7 bottoms from 48 standard K-lines, while Gemini's results included a similar number of identified patterns, showcasing both models' capabilities in this area [23][30] - The evaluation highlights differences in the models' approaches to defining and processing K-line relationships, with Gemini's methodology being more rigorous in certain aspects [40]