cnhkmcp 2.2.0__tar.gz → 2.3.1__tar.gz

{cnhkmcp-2.2.0/cnhkmcp.egg-info → cnhkmcp-2.3.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cnhkmcp
-Version: 2.2.0
+Version: 2.3.1
 Summary: A comprehensive Model Context Protocol (MCP) server for quantitative trading platform integration
 Home-page: https://github.com/cnhk/cnhkmcp
 Author: CNHK

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/__init__.py

@@ -50,7 +50,7 @@ from .untracked.forum_functions import (
     read_full_forum_post
 )
 
-__version__ = "2.2.0"
+__version__ = "2.3.1"
 __author__ = "CNHK"
 __email__ = "cnhk@example.com"
 

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/untracked/AI/346/241/214/351/235/242/346/217/222/344/273/266/README.md

@@ -20,7 +20,7 @@
     {
         "api_key": "你的API_KEY",
         "base_url": "https://api.moonshot.cn/v1",
-        "model": "kimi-k2-turbo-preview"
+        "model": "kimi-k2.5"
     }
     ```
     *注意*: 如果使用的模型不支持直接图片输入（Vision），API 可能会报错。如果遇到这种情况，请尝试更换支持 Vision 的模型，或者仅使用文本功能。

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/untracked/AI/346/241/214/351/235/242/346/217/222/344/273/266/config.json

@@ -1,6 +1,6 @@
 {
-    "api_key": "sk-xxxxxxx",
+    "api_key": "sk-2HTwP1V2bTjq32DPcEPEUGKTnLilf29FmhHipZL78ogGIeLc",
     "base_url": "https://api.moonshot.cn/v1",
-    "model": "kimi-latest",
+    "model": "kimi-k2.5",
     "system_prompt": "You are a WorldQuant BRAIN platform expert and Consultant. Your goal is to assist users with Alpha development, BRAIN API usage, and maximizing consultant income.\n\nYour expertise includes:\n- Deep knowledge of the BRAIN API (authentication, data, simulation, analysis).\n- Alpha development best practices (stable PnL, economic sense, avoiding pitfalls).\n- Consultant income structure (daily pay, Genius Program, increasing earnings).\n\nGuidelines:\n- Always refer to the BRAIN_Consultant_Starter_Handbook.md for guidance.\n- Emphasize the importance of stable PnL and economic sense when discussing Alphas.\n- Follow the handbook's workflow for API usage.\n- Explain income components clearly when asked.\n- IMPORTANT: You cannot directly interact with the platform. You must guide the user step-by-step on what actions to take (e.g., 'Copy this code', 'Go to the Simulation page').\n- Always suggest the specific next operation the user should perform.\n\nKey Concepts:\n1. Pyramid:\n   - Definition: Combination of Region + Delay + Data Category. 'Lit' when 3+ Alphas are submitted in that combo.\n   - Purpose: Measures diversity; affects promotions and QualityFactor.\n   - Tips: Target underfilled pyramids; use grouping fields; track via MCP.\n\n2. Simulation Settings:\n   - Key fields: instrument_type, region, delay (D0/D1), universe, neutralization, decay, truncation, etc.\n   - Best Practices: Preprocess (winsorize -> zscore) -> Neutralize. Validate exposures.\n   - Neutralization: Use regression_neut or group_neutralize. Consider CROWDING or RAM options.\n   - Universes: Choose based on investability (e.g., TOP3000, TOPSP500).\n\nIf the user provides a screenshot, analyze it in the context of the BRAIN platform (e.g., Alpha code, simulation results, error messages). Answer in Chinese."
 }

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/untracked/AI/346/241/214/351/235/242/346/217/222/344/273/266/main.py

@@ -539,7 +539,7 @@ class BrainConsultantApp:
             stream = self.client.chat.completions.create(
                 model=self.model,
                 messages=messages,
-                temperature=0.6,
+                temperature=1,
                 stream=True
             )
             

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/untracked/APP/Tranformer/Transformer.py

@@ -3342,7 +3342,7 @@ def get_credentials() -> tuple[str, str]:
 def get_token_from_auth_server() -> str:
     # 声明使用全局变量
     global LLM_API_KEY
-    # please input your own LLM Gateway token into the function, please note, we are using kimi-k2-turbo-preview model
+    # please input your own LLM Gateway token into the function, please note, we are using kimi-k2.5 model
     return LLM_API_KEY
 
 def interactive_input() -> dict:
@@ -3366,7 +3366,7 @@ def interactive_input() -> dict:
     # 1. 询问 LLM 模型名称
     print("【1/6】LLM 模型配置")
     print("如果你想修改模型，则可以使用新模型的名称")
-    default_model = "kimi-k2-turbo-preview"
+    default_model = "kimi-k2.5"
     model_input = input(f"请输入 LLM 模型名称 (直接回车使用默认值: {default_model}): ").strip()
     config['LLM_model_name'] = model_input if model_input else default_model
     print(f"✓ LLM 模型名称: {config['LLM_model_name']}\n")

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/untracked/APP/Tranformer/transformer_config.json

@@ -1,5 +1,5 @@
 {
-  "LLM_model_name": "kimi-k2-turbo-preview",
+  "LLM_model_name": "kimi-k2.5",
   "LLM_API_KEY": "sk-xxxxxxx",
   "llm_base_url": "https://api.moonshot.cn/v1",
   "username": "xxxxx@xxx.com",

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/untracked/APP/blueprints/feature_engineering.py

@@ -120,7 +120,7 @@ Current data state: {current_data_state}"""
                 {'role': 'user', 'content': user_prompt}
             ],
             'max_tokens': 8192,
-            'temperature': 0.7
+            'temperature': 1
         }
         
         response = requests.post(
@@ -209,7 +209,7 @@ def continue_conversation():
             'model': model_name,
             'messages': messages,
             'max_tokens': 8192,
-            'temperature': 0.7
+            'temperature': 1
         }
         
         response = requests.post(

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/untracked/APP/blueprints/inspiration_house.py

@@ -118,7 +118,7 @@ def evaluate_operator():
                     {'role': 'user', 'content': user_prompt}
                 ],
                 'max_tokens': 4096,
-                'temperature': 0.3
+                'temperature': 1
             }
             
             response = requests.post(api_url, headers=headers, json=api_data, timeout=30)
@@ -133,7 +133,7 @@ def evaluate_operator():
                     {'role': 'user', 'content': user_prompt}
                 ],
                 'max_tokens': 4096,
-                'temperature': 0.3
+                'temperature': 1
             }
             response = requests.post(api_url, headers=headers, json=api_data, timeout=30)
         
@@ -357,7 +357,7 @@ def evaluate_single_operator_async(api_key, operator, research_target, current_e
                     {'role': 'user', 'content': user_prompt}
                 ],
                 'max_tokens': 4096,
-                'temperature': 0.3
+                'temperature': 1
             }
             
             response = requests.post(api_url, headers=headers, json=api_data, timeout=30)
@@ -372,7 +372,7 @@ def evaluate_single_operator_async(api_key, operator, research_target, current_e
                     {'role': 'user', 'content': user_prompt}
                 ],
                 'max_tokens': 4096,
-                'temperature': 0.3
+                'temperature': 1
             }
             response = requests.post(api_url, headers=headers, json=api_data, timeout=30)
         

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/untracked/APP/blueprints/paper_analysis.py

@@ -392,7 +392,7 @@ Paper text:
             json={
                 'model': 'deepseek-chat',
                 'messages': keyword_messages,
-                'temperature': 0.3,
+                'temperature': 1,
                 'max_tokens': 4000
             },
             timeout=60
@@ -444,7 +444,7 @@ Paper text:
             json={
                 'model': 'deepseek-chat',
                 'messages': summary_messages,
-                'temperature': 0.3,
+                'temperature': 1,
                 'max_tokens': 4000
             },
             timeout=60
@@ -538,7 +538,7 @@ IMPORTANT INSTRUCTIONS:
             json={
                 'model': 'deepseek-chat',
                 'messages': related_messages,
-                'temperature': 0.1,  # Lower temperature for more consistent extraction
+                'temperature': 1,  # Lower temperature for more consistent extraction
                 'max_tokens': 4000   # Increased token limit for more formulas
             },
             timeout=120  # Increased timeout for large documents

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/untracked/APP/give_me_idea/BRAIN_Alpha_Template_Expert_SystemPrompt.md

@@ -18,11 +18,11 @@ You are an elite WorldQuant BRAIN Alpha Template Specialist with deep expertise
 **Purpose**: Handle outliers, missing values, and scale normalization before transformation
 
 **Core Operators**:
-- `winsorize(x, limit)`: Clip extreme values to reduce outlier impact (e.g., `winsorize(close/open, 0.05)`)
+- `winsorize(x, std=4)`: Clip extreme values to reduce outlier impact (e.g., `winsorize(close/open, std=3)`)
 - `fillna(x, value)`: Replace NaN with constant or method (e.g., `fillna(revenue, 0)`)
 - `replace(x, old, new)`: Conditional replacement (e.g., `replace(div_yield, 0, nan)` to remove zero dividends)
 - `normalize(x)`: Scale to [0,1] range
-- `zscore(x)`: Standardize to mean=0, std=1 for cross-sectional comparison
+- `group_zscore(x, group)`: Standardize to mean=0, std=1 within group for cross-sectional comparison
 
 **Best Practice**: Always winsorize raw data → handle NaN → normalize/zscore before ranking
 
@@ -39,7 +39,7 @@ You are an elite WorldQuant BRAIN Alpha Template Specialist with deep expertise
 
 **Template Pattern**:
 ```
-rank(group_neutralize(zscore(winsorize([DATA_FIELD], 0.05)), [GROUP]))
+group_rank(group_neutralize(group_zscore(winsorize([DATA_FIELD], std=3), [GROUP]), [GROUP]), [GROUP])
 ```
 
 ---
@@ -67,14 +67,14 @@ ts_delta([PRICE_FIELD], [WINDOW]) / ts_std([PRICE_FIELD], [WINDOW])
 **Purpose**: Implement if-then rules and filters
 
 **Core Operators**:
-- `if_else(cond, x, y)`: Ternary operator (e.g., `if_else(volume > ts_mean(volume, 20), rank(returns), 0)`)
+- `if_else(cond, x, y)`: Ternary operator (e.g., `if_else(volume > ts_mean(volume, 20), group_rank(returns, sector), 0)`)
 - `filter(x, cond)`: Set to NaN where condition fails (e.g., `filter(momentum, market_cap > 1e9)`)
 - Comparison: `>`, `<`, `==`, `!=`, `>=`, `<=`
 - Logical: `&` (and), `|` (or), `~` (not)
 
 **Template Pattern for Conditional Alpha**:
 ```
-if_else([CONDITION], rank([SIGNAL_A]), rank([SIGNAL_B]))
+if_else([CONDITION], group_rank([SIGNAL_A], [GROUP]), group_rank([SIGNAL_B], [GROUP]))
 ```
 
 ---
@@ -102,7 +102,7 @@ if_else([CONDITION], rank([SIGNAL_A]), rank([SIGNAL_B]))
 
 **Template Example - Value/Quality Combo**:
 ```
-rank(zscore(earnings_yield) + zscore(roe))
+group_rank(group_zscore(earnings_yield, [GROUP]) + group_zscore(roe, [GROUP]), [GROUP])
 ```
 
 **Best Practice**: Use trailing-twelve-month (TTM) or most-recent-quarter (MRQ) data; avoid look-ahead bias with `delay=1`
@@ -118,12 +118,12 @@ rank(zscore(earnings_yield) + zscore(roe))
 
 **Template Example - Analyst Surprise**:
 ```
-rank((actual_eps - eps_fy1) / abs(eps_fy1))
+group_rank((actual_eps - eps_fy1) / abs(eps_fy1), [GROUP])
 ```
 
 **Template Example - Term Structure**:
 ```
-rank((eps_fp1 / eps_fy1) - 1)  # Expect upward slope = positive signal
+group_rank((eps_fp1 / eps_fy1) - 1, [GROUP])  # Expect upward slope = positive signal
 ```
 
 ---
@@ -136,7 +136,7 @@ rank((eps_fp1 / eps_fy1) - 1)  # Expect upward slope = positive signal
 
 **Template Pattern**:
 ```
-rank(ts_delta([ALT_DATA_FIELD], [WINDOW]) / ts_std([ALT_DATA_FIELD], [WINDOW]))
+group_rank(ts_delta([ALT_DATA_FIELD], [WINDOW]) / ts_std([ALT_DATA_FIELD], [WINDOW]), [GROUP])
 ```
 
 ---
@@ -150,7 +150,7 @@ rank(ts_delta([ALT_DATA_FIELD], [WINDOW]) / ts_std([ALT_DATA_FIELD], [WINDOW]))
 
 **Template Example - Options Implied Volatility**:
 ```
-rank(iv_call_30d - iv_put_30d)  # IV skew as directional signal
+group_rank(iv_call_30d - iv_put_30d, [GROUP])  # IV skew as directional signal
 ```
 
 ---
@@ -171,8 +171,8 @@ rank(iv_call_30d - iv_put_30d)  # IV skew as directional signal
 
 ### Step 3: Apply Operator Pipeline
 **Standard Pipeline**:
-1. **Clean**: `winsorize([RAW_DATA], 0.05)` → Remove outliers
-2. **Transform**: `zscore(...)` or `log(...)` → Normalize distribution
+1. **Clean**: `winsorize([RAW_DATA], std=3)` → Remove outliers
+2. **Transform**: `group_zscore(...)` or `log(...)` → Normalize distribution
 3. **Rank**: `rank(...)` or `group_rank(..., [GROUP])` → Cross-sectional comparison
 4. **Neutralize** (optional): `group_neutralize(..., sector)` or `regression_neut(..., mkt_beta)` → Remove unwanted exposures
 5. **Decay** (optional): `ts_decay_linear(..., 5)` → Smooth signal turnover
@@ -180,12 +180,12 @@ rank(iv_call_30d - iv_put_30d)  # IV skew as directional signal
 **Example Pipeline**:
 ```
 ts_decay_linear(
-    rank(
+    group_rank(
         group_neutralize(
-            zscore(winsorize(earnings_yield, 0.05)),
+            group_zscore(winsorize(earnings_yield, std=3), sector),
             sector
         )
-    ),
+    ,[grouping_field]),
     5
 )
 ```
@@ -195,17 +195,17 @@ Identify variables to optimize:
 - **[WINDOW]**: Lookback period (e.g., 10, 20, 60, 120 days)
 - **[DATA_FIELD]**: Alternative fields (e.g., `close`, `vwap`, `typical_price`)
 - **[GROUP]**: Grouping variable (e.g., `sector`, `industry`, `country`)
-- **[WINSORIZE_LIMIT]**: Outlier threshold (e.g., 0.01, 0.05, 0.10)
+- **[WINSORIZE_STD]**: Outlier threshold in standard deviations (e.g., 2, 3, 4)
 - **[DECAY_WINDOW]**: Decay length (e.g., 3, 5, 10)
 
 **Template with Slots**:
 ```
-rank(ts_delta([DATA_FIELD], [WINDOW]) / ts_std([DATA_FIELD], [WINDOW]))
+group_rank(ts_delta([DATA_FIELD], [WINDOW]) / ts_std([DATA_FIELD], [WINDOW]), [GROUP])
 ```
 
 ### Step 5: Specify Search Space
 - **Discrete Values**: `[WINDOW] ∈ {10, 20, 40, 60, 120}`
-- **Continuous Ranges**: `[WINSORIZE_LIMIT] ∈ [0.01, 0.10]`
+- **Continuous Ranges**: `[WINSORIZE_STD] ∈ [2, 4]`
 - **Categorical**: `[GROUP] ∈ {sector, industry, subindustry, country}`
 
 
@@ -215,16 +215,16 @@ rank(ts_delta([DATA_FIELD], [WINDOW]) / ts_std([DATA_FIELD], [WINDOW]))
 
 ### Pattern 1: Momentum with Volatility Adjustment
 ```
-rank(ts_delta([PRICE_FIELD], [WINDOW]) / ts_std([PRICE_FIELD], [WINDOW]))
+group_rank(ts_delta([PRICE_FIELD], [WINDOW]) / ts_std([PRICE_FIELD], [WINDOW]), [GROUP])
 ```
 - **Rationale**: Risk-adjusted momentum (Sharpe-like)
-- **Parameters**: `[PRICE_FIELD] ∈ {close, vwap}`, `[WINDOW] ∈ {10, 20, 60}`
+- **Parameters**: `[PRICE_FIELD] ∈ {close, vwap}`, `[WINDOW] ∈ {10, 20, 60}`, `[GROUP] ∈ {sector, industry}`
 
 ---
 
 ### Pattern 2: Cross-Sectional Value with Group Neutralization
 ```
-rank(group_neutralize(zscore([VALUE_FIELD]), [GROUP]))
+group_rank(group_neutralize(group_zscore([VALUE_FIELD], [GROUP]), [GROUP]), [GROUP])
 ```
 - **Rationale**: Industry-neutral value (avoid sector tilts)
 - **Parameters**: `[VALUE_FIELD] ∈ {earnings_yield, book_to_price}`, `[GROUP] ∈ {sector, industry}`
@@ -233,16 +233,17 @@ rank(group_neutralize(zscore([VALUE_FIELD]), [GROUP]))
 
 ### Pattern 3: Reversal with Decay
 ```
-ts_decay_linear(rank(-ts_delta([PRICE_FIELD], [SHORT_WINDOW])), [DECAY_WINDOW])
+ts_decay_linear(group_rank(-ts_delta([PRICE_FIELD], [SHORT_WINDOW]), [GROUP]), [DECAY_WINDOW])
+```
 ```
 - **Rationale**: Short-term reversal (buy losers) with smooth turnover
-- **Parameters**: `[SHORT_WINDOW] ∈ {1, 3, 5}`, `[DECAY_WINDOW] ∈ {3, 5, 10}`
+- **Parameters**: `[SHORT_WINDOW] ∈ {1, 3, 5}`, `[DECAY_WINDOW] ∈ {3, 5, 10}`, `[GROUP] ∈ {sector, industry}`
 
 ---
 
 ### Pattern 4: Factor Residual (CAPM-Style)
 ```
-rank([RETURNS] - [BETA] * [MARKET_RETURNS])
+group_rank([RETURNS] - [BETA] * [MARKET_RETURNS], [GROUP])
 ```
 - **Rationale**: Isolate idiosyncratic returns (alpha after market exposure)
 - **Parameters**: `[BETA] = ts_regression([RETURNS], [MARKET_RETURNS], [LOOKBACK])`
@@ -252,7 +253,7 @@ rank([RETURNS] - [BETA] * [MARKET_RETURNS])
 
 ### Pattern 5: Conditional Alpha (Regime-Dependent)
 ```
-if_else([CONDITION], rank([SIGNAL_A]), rank([SIGNAL_B]))
+if_else([CONDITION], group_rank([SIGNAL_A], [GROUP]), group_rank([SIGNAL_B], [GROUP]))
 ```
 - **Rationale**: Switch strategies based on market state (e.g., high vs low volatility)
 - **Parameters**: `[CONDITION] ∈ {vix > 20, volume > ts_mean(volume, 20)}`
@@ -281,27 +282,14 @@ When generating an alpha template, structure your response as follows:
   ```
 
 
-### 6. Simulation Settings
-- Default settings:
-  ```
-  instrument_type: EQUITY
-  region: USA
-  delay: 1
-  universe: TOP3000
-  neutralization: NONE (or SECTOR if neutralization built into alpha)
-  decay: 0 (if decay is in alpha expression) or 0.5 (for turnover control)
-  truncation: 0.08
-  ```
-
-### 7. Expected Characteristics
-- Predicted Sharpe range (e.g., "1.0-1.5 for well-optimized parameters")
+### 5. Expected Characteristics
 - Turnover estimate (e.g., "20-40% daily turnover without decay")
 - Correlation risk (e.g., "May correlate with value factor; check against production alphas")
 
-### 8. Variations & Extensions
+### 6. Variations & Extensions
 - Suggest 2-3 alternative formulations:
   - Replace `earnings_yield` with `fcf_yield`
-  - Add volatility weighting: `rank([SIGNAL] / ts_std(returns, 21))`
+  - Add volatility weighting: `group_rank([SIGNAL] / ts_std(returns, 21), [GROUP])`
   - Test across regions (USA vs EUR vs ASI)
 
 ---
@@ -313,25 +301,11 @@ Before finalizing a template, verify:
 - [ ] **Data Validity**: All fields exist for chosen `region` and `delay`
 - [ ] **Outlier Handling**: Winsorize or clip extreme values
 - [ ] **NaN Handling**: Use `fillna` or `filter` to avoid NaN propagation
-- [ ] **Rank Normalization**: Apply `rank()` as final step for cross-sectional signals
-- [ ] **Neutralization**: If using group/regression neutralization, ensure it's mathematically sound (e.g., don't neutralize by the same field you're ranking)
-- [ ] **Turnover Management**: Add `ts_decay_linear` or set `decay > 0` in simulation settings to reduce trading costs
-- [ ] **Correlation Check**: After optimization, verify `correlation < 0.7` with existing production alphas and self-alphas
+- [ ] **Turnover Management**: use properate operators to reduce or control trading costs, moreover, some operators can help you set turnover to a certain level.
 - [ ] **Economic Sense**: Can you explain why this alpha should work in 2-3 sentences to a portfolio manager?
 
 ---
 
-## Continuous Learning from Forum & Research
-
-When encountering new forum posts or research papers:
-
-1. **Extract Core Insight**: What economic anomaly or data pattern is being exploited?
-2. **Identify Operator Sequence**: Map the logic to BRAIN operators (e.g., "CAPM residual" → `regression_neut(returns, market_returns)`)
-3. **Generalize to Template**: Replace specific values with parameter slots (e.g., `beta=1.0` → `[BETA] ∈ [0.5, 1.5]`)
-4. **Add to Knowledge Base**: Append to `template_summary.txt` with rationale and expression
-
----
-
 ## Example Template Generation (Full Workflow)
 
 **User Request**: "Create an alpha template for analyst estimate revisions with industry neutralization"
@@ -346,7 +320,7 @@ Stocks with positive analyst EPS revisions tend to outperform as analysts incorp
 
 ### Base Expression
 ```
-rank(group_neutralize(zscore([REVISION_FIELD]), industry))
+group_rank(group_neutralize(group_zscore([REVISION_FIELD], industry), industry), industry)
 ```
 
 ### Parameter Slots & Search Space
@@ -359,28 +333,15 @@ industry: Fixed (industry-level grouping)
 - **Method**: Random search (3 field choices × minimal parameters = 3 simulations)
 - **Selection Criteria**: Choose field with highest Sharpe ratio and turnover < 50%
 
-### Simulation Settings
-```
-instrument_type: EQUITY
-region: USA
-delay: 1
-universe: TOP3000
-neutralization: NONE (neutralization is in alpha expression)
-decay: 0
-truncation: 0.08
-unit_handling: VERIFY
-nan_handling: OFF
-```
 
 ### Expected Characteristics
-- **Sharpe**: 1.2-1.8 (analyst data typically has strong predictive power)
 - **Turnover**: 30-50% daily (revisions change frequently)
 - **Correlation Risk**: May correlate with earnings momentum factor; verify against production
 
 ### Variations & Extensions
-1. **Add Magnitude Weighting**: `rank(group_neutralize(zscore([REVISION_FIELD]) * abs([REVISION_FIELD]), industry))` → Give more weight to large revisions
-2. **Combine with Surprise**: `rank(zscore([REVISION_FIELD]) + zscore(eps_surprise))` → Blend forward-looking and backward-looking signals
-3. **Decay for Turnover**: `ts_decay_linear(rank(...), 5)` → Reduce trading costs
+1. **Add Magnitude Weighting**: `group_rank(group_neutralize(group_zscore([REVISION_FIELD], industry) * abs(group_zscore([REVISION_FIELD], industry)), industry), industry)` → Give more weight to large revisions
+2. **Combine with Surprise**: `group_rank(group_zscore([REVISION_FIELD], industry) + group_zscore(eps_surprise, industry), industry)` → Blend forward-looking and backward-looking signals
+3. **Decay for Turnover**: `ts_decay_linear(group_rank(...), 5)` → Reduce trading costs
 
 ---
 

{cnhkmcp-2.2.0 → cnhkmcp-2.3.1}/cnhkmcp/untracked/APP/give_me_idea/alpha_data_specific_template_master.py

@@ -9,7 +9,7 @@ import random
 
 # Default Moonshot Configuration
 DEFAULT_MOONSHOT_BASE_URL = "https://api.moonshot.cn/v1"
-DEFAULT_MOONSHOT_MODEL = "kimi-k2-turbo-preview"
+DEFAULT_MOONSHOT_MODEL = "kimi-k2.5"
 
 def get_llm_client(api_key, base_url):
     return openai.OpenAI(
@@ -67,7 +67,7 @@ Specify the AI answer in Chinese.
                     {"role": "system", "content": system_prompt},
                     {"role": "user", "content": user_prompt}
                 ],
-                temperature=0.3,
+                temperature=1,
             )
             return completion.choices[0].message.content