AI-InvestiBot

AI-InvestiBot provides scripts and utilities for experimenting with indicator-heavy LSTM models, walk-forward validation, and automated decision logic across equities. The repository includes helper modules for gathering data, crafting indicators, training models, and connecting them to decision tooling.

Contents

• Highlights
• Roadmap
• Quick Start
• Training Workflow
  • Data Retrieval and Caching
• Feature Engineering
• Model Zoo
• Hold-out Monitoring
• Decision Automation
• Support

---

Highlights

• Indicator-centric cache: get_info.py produces Stocks/<TICKER>/info.json with earnings diffs, kumo clouds, supertrends, and other custom indicators used by the LSTM pipelines.
• Pluggable model builders: PriceModel, PercentageModel, and DirectionalModel accept create_model=... callbacks, so you can prototype new architectures without rewriting the training loop.
• Walk-forward workflow: train(test=True) hides the boilerplate for hold-out splits and scaler tracking. model.test() reuses the configuration for a fast sanity check.
• Batch experimentation helpers: holdout_monitor.py accepts a JSON sample describing experiments and emits JSON or CSV summaries to compare indicator mixes or model families.
• Prototype decision stack: decision_maker.py loads multiple saved strategies, refreshes their cached indicators, and feeds a DecisionTreeClassifier for a trade vote.
• Serverless skeleton: implementation.py includes a Lambda-ready loop for anyone who wants to run lightweight decision logic without a dedicated workstation.

Roadmap

• Hold-out driven accuracy targets — ✅
• Callback-based training API — ✅
• More validation/tests — 🔄
• PercentageModel follow-ups — 🔄

Quick Start

⚠️ Real-time trading remains experimental. Use the tooling for research/backtesting until further notice.

1. Install dependencies: pip install -r requirements.txt
2. Configure secrets: Copy secrets_example.config → secrets.config, add API keys (trading, AlphaVantage fallback, etc.).
3. Cache data: run python -m AI-InvestiBot.get_info with the symbol flag set to AAPL. The helper fetches Yahoo Finance history (with AlphaVantage or Stooq fallbacks) and writes Stocks/<SYMBOL>/info.json.
4. Train a model:

   from AI-InvestiBot.models import PriceModel
   from AI-InvestiBot.custom_objects import create_lightweight_model
   
   model = PriceModel(stock_symbol="AAPL", information_keys=["Close","returns_zscore"])
   model.train(create_model=create_lightweight_model, epochs=300, patience=10, test=True)
   model.save("Stocks/AAPL/MyPriceModel")

5. Evaluate hold-out:

   directional, spatial, rmse, rmsse, homogenous = model.test(show_graph=False)

6. Batch comparisons (optional): run python -m AI-InvestiBot.holdout_monitor with your experiment config path and desired output location to produce JSON and CSV summaries.

Keep Stocks/ out of version control—models and cached data regenerate locally.

Training Workflow

Data Retrieval and Caching

• Download historical OHLCV — trading_funcs.download_stock_history. Retries Yahoo and falls back to AlphaVantage or Stooq when rate-limited.
• Derive indicators and metadata — get_info.py. Produces info.json, dynamic_tuning.json, plus scaler references per symbol.
• Inspect caches — validate_inputs.py. Spot-checks indicator quality, counts samples, and flags missing or invalid entries.

Each indicator drawer writes to JSON so the training loop can load features with get_relavant_values() and feed them into the neural network builders.

Feature Engineering

Key context-aware indicators you can toggle via information_keys:

• returns_zscore: Rolling 20-day z-score of daily returns.
• volatility_14: Two-week realized volatility for regime awareness.
• trend_strength: Relative gap between 50/200-day EMAs.
• ema_spread_10_40, volume_surge, atr_14: Derived in trading_funcs.
• earnings_flag: Binary flag spanning a ±3-day window around earnings.

Non-daily series (earnings dates/diffs) are aligned alongside daily bars for richer context.

Model Zoo

• PriceModel: Predicts price directly after scaling against historical min/max.
• PercentageModel: Predicts percentage returns over sliding windows.
• DirectionalModel: Optimized for sign accuracy, with curriculum stages, balanced focal loss, and automatic threshold calibration.
• Custom builders:
  • create_lightweight_model: Single LSTM layer for rapid prototyping.
  • create_context_gated_model: Adds a context gate derived from global averages.
  • create_probabilistic_model: Predicts mean + log-variance (heteroscedastic loss).
  • create_directional_model / create_directional_model_focal: Sigmoid outputs backed by balanced focal loss.

The training loop accepts callbacks, so you can plug in any Keras Model factory that matches the expected input shape.

Hold-out Monitoring

AI-InvestiBot/holdout_monitor.py packages the walk-forward pipeline into a CLI:

Run python -m AI-InvestiBot.holdout_monitor with configs/experiments.json and an output file path. The command reads experiments, executes each training run, and writes its findings to JSON and CSV.

• Reads { "experiments": [ { "stock_symbol": "...", "information_keys": [...], "create_model": "gated", ... } ] }
• Runs PriceModel.train(test=True) per entry and records directional/spatial/RMSE metrics for the hold-out slice.
• Writes both JSON and CSV summaries for quick spreadsheet analysis.

Decision Automation

decision_maker.py loads multiple trained strategy models (Impulse MACD, breakout, RSI, supertrends, etc.), builds cached indicator windows (online or offline), and feeds predictions into a DecisionTreeClassifier.

Features:

• Offline cache reader for disconnected environments (Stocks/<SYMBOL>/info.json).
• Automatic re-computation of indicators per 14-day step when running forward in time.
• Optional TARGET_SYMBOLS override to focus the evaluation.
• Designed to slot into the ResourceManager flows so you can allocate capital based on aggregated votes rather than a single model.

Support

• Discord: https://dsc.gg/ai-investibot/ (custom vanity link)
• Issues/PRs: Please open tickets for bugs, docs gaps, or feature proposals. Contributions should avoid checking in regenerated Stocks/ assets.

---

Thanks for taking a look at AI-InvestiBot. If you build something useful on top of the framework, feel free to share it through an issue or pull request so others can benefit as well.