AlphaSpire

AI agent-driven fully automated Alpha mining.

---

⚠️ Attention

Please note that this tool only performs backtest simulations on the platform (Simulate) and does not execute any submission actions (Submit). Excessive simulations without valid submissions may result in your account being locked. Please make sure to limit the number of backtests when using this tool.

Work Pipeline

                ┌────────────────────────────┐
                │   Data Acquisition Module  │
                │ (Scraper & Local Storage)  │
                └─────────────┬──────────────┘
                              │
                ┌─────────────▼──────────────┐
                │   Preprocessing & Parsing  │
                │ (Text cleaning, Metadata)  │
                └─────────────┬──────────────┘
                              │
                ┌─────────────▼──────────────┐
                │     Researcher (LLM)       │
                │ (LangChain / Prompting)    │
                └─────────────┬──────────────┘
                              │
                ┌─────────────▼──────────────┐
                │  Alpha Template Generating │
                │ (Factor/Signal Extraction) │
                └─────────────┬──────────────┘
                              │
                ┌─────────────▼──────────────┐
                │          Evaluator         │
                │ (Submit to WQ or local sim)│
                └─────────────┬──────────────┘
                              │
                ┌─────────────▼──────────────┐
                │    Results Collecting      │
                │ (Store Scores, Rank Alphas)│
                └─────────────┬──────────────┘
                              │
                ┌─────────────▼──────────────┐
                │Results analyzing &Iteration│
                │                            │
                └────────────────────────────┘

Architecture

data

The data folder stores posts, operators, fields, operator types, field types, generated alpha templates, generated alphas, and backtesting results information.

prompts

Stores prompt templates used in workflows.

utils

1. Get the operator set of Fast Expression and save it to ./data/wq_operators
2. Get the field collection of Fast Expression and save it to ./data/wq_fields
3. Manual rules classify operators and generate operator types required by templates. (save to ./data/wq_template_operators)
4. Clustering and large language models are used to assist in generating the field types used in templates. (save to ./data/wq_template_fields)
5. Various text processing and other miscellaneous.

scraper

1. Scrape posts from the WorldQuant forum and store them in ./data/wq_posts/raw_posts.
2. Extract the main information from the original html text of the post into a json file and store it in ./data/wq_posts/processed_posts
3. Use a large language model to determine whether the post has the potential to generate alpha, and if so, save it to ./data/wq_posts/helpful_posts

researcher

1. Use LLM to summarize hypothesis from blog and save it to ./data/hypothesis_db_v2
2. Use LLM to generate template from hypothesis and save it to ./data/template_db_v2
3. Multiple loops generate alphas based on template, field type, and operator type and save them to ./data/alpha_db_v2/all_alphas

evaluator

Use the WorldQuant backtesting API to evaluate alpha performance and save the results to data/alpha_db_v2/backtest_result

Deployment

1. Create a conda environment

   conda env create -f environment.yml
   conda activate alphaspire

2. Fill in the configuration file (config.yaml)

   # ===============================
   # Global Configuration File
   # ===============================
   
   # --- OpenAI API Settings ---
   openai_base_url: "todo" # e.g. https://api.deepseek.com
   openai_api_key: "todo" # e.g. sk-...
   openai_model_name: "todo" # e.g. deepseek-chat
   reasoner_model_name: "todo" # e.g. deepseek-reasoner
   
   # --- WorldQuant Platform Credentials ---
   worldquant_account: "todo"
   worldquant_password: "todo"
   
   worldquant_login_url: "https://platform.worldquantbrain.com/sign-in"
   worldquant_api_auth: "https://api.worldquantbrain.com/authentication"
   worldquant_consultant_posts_url: "https://support.worldquantbrain.com/hc/en-us/community/topics/18910956638743-顾问专属中文论坛"
   # You can also choose any other WorldQuant Forum URL you have access to.
   
   # --- Dataset from WorldQuant Brain
   enabled_field_datasets: # Select the field database you want to use to build alphas.
     - pv1                 # Database name reference ./data/wq_fields
     - fundamental6
     - analyst4
     - model16
     - news12   

3. One-click operation

   • Full process operation

      python3 main.py

   • Only run post crawl

      python3 main_scraper.py

   • Only run Alpha generation

      python3 main_researcher.py

   • Only run Alpha backtests

      ./test_script.sh

     OR

      python3 main_evaluator.py

Notice

Regarding WorldQuant backtest parameter settings, currently only manual modification of the parameters in the following code section in evaluator/backtest_with_wq_mul.py is supported. These will be moved to a separate config file in future versions.

payload = {
    "type": "REGULAR",
    "settings": {
        "instrumentType": "EQUITY",
        "region": "ASI",
        "universe": "MINVOL1M",
        "delay": 1,
        "decay": 6,
        "neutralization": "SUBINDUSTRY",
        "truncation": 0.01,
        "pasteurization": "ON",
        "unitHandling": "VERIFY",
        "nanHandling": "ON",
        "maxTrade": "ON",
        "language": "FASTEXPR",
        "visualization": False,
    },
    "regular": fixed_expr
}