I Ching Stock Prediction Model

Python implementation of the paper:

"Incorporating I Ching Knowledge Into Prediction Task via Data Mining" Liu et al., Journal of Database Management, Volume 34, Issue 3 DOI: 10.4018/JDM.322097

The model combines traditional I Ching divination with modern machine learning to predict stock price trends (Positive / Neutral / Negative). It outperforms SVM, XGBoost, RF, GRU, and LSTM-Att in the paper's experiments (Macro F1: 63.33% vs. best baseline 61.08%).

All Features
    │
    ▼
Random Forest (Gini importance)
    │  top-30 features
    ▼
Three Vitals Mapping
  Ten (天) — macro/environment
  Chi (地) — size & flow
  Jin (人) — fundamentals
    │  top-2 per Vital = 6 features
    ▼
┌─────────────────────────────┐       ┌──────────────────────┐
│     Hexagram Generator      │       │       Decoder        │
│                             │       │                      │
│  Four Signs (sliding k=8)   │──────▶│  Hexagram Explainer  │
│  LaoYin=6 / ShaoYin=8      │       │  (Algorithm Two)     │
│  ShaoYang=7 / LaoYang=9    │       │                      │
│         ↓                  │       │  Sentiment Classifier │
│  6-bit Original Hexagram   │       └──────────────────────┘
│  6-bit Changed Hexagram    │               │
└─────────────────────────────┘               ▼
                                    Positive / Neutral / Negative

I Ching's framework for categorising all influencing factors:

Each of the 6 feature values is mapped to a Four Signs code via a sliding window of size 8:

Where A = (1/8) × range, B = (3/8) × range.

• 6 Yao values (0/1) form a 6-bit binary string (bottom to top)
• The Vital ordering (which Vital maps to which Yao positions) is selected by trying all 6 permutations of [Ten, Jin, Chi] during fit() and keeping the one with the best macro F1
• Original hexagram: raw Yao values
• Changed hexagram: LaoYang (9) flips 1→0, LaoYin (6) flips 0→1

Selects the I Ching explanation based on the number of moving Yaos (YbNum):

Huber regression is fit over a rolling window of closing prices. The slope coefficient a is classified using mean ± std thresholds derived from the normal distribution of all slopes across the dataset:

• Positive (rising): a > mean + std
• Neutral (stable): mean − std ≤ a ≤ mean + std (~68% of samples)
• Negative (falling): a < mean − std

This matches the paper's statement: "the distribution of a is counted, and finally the label category is obtained according to the distribution interval of a."

Huber loss is used for robustness to outliers:

L(y, f(x)) = 0.5*(y-f(x))²          if |y-f(x)| ≤ δ
             δ*(|y-f(x)| - 0.5*δ)   otherwise

iching/
├── iching_data.py                   # 64 hexagram dict, 384 Yao explanations, sentiment scores
├── iching_wilhelm_translation.js    # Local copy of adamblvck/iching-wilhelm-dataset (MIT)
├── import_wilhelm.py                # One-shot importer: regenerates iching_data.py from dataset
├── label_construction.py            # Huber regression → {Positive, Neutral, Negative} labels
├── feature_selection.py             # RF Gini importance + Three Vitals mapping → 6 features
├── hexagram_generator.py            # Algorithm One: Four Signs + hexagram generation
├── hexagram_decoder.py              # Algorithm Two: Hexagram Explainer + classifier
├── iching_model.py                  # Algorithm Three: full pipeline (IChingModel class)
├── baselines.py                     # SVM, XGBoost, RF, KNN, GRU, LSTM-Att models
├── benchmark.py                     # Train all 7 models and print comparison table
└── demo.py                          # End-to-end demo (AAPL via yfinance or synthetic data)

pip install numpy pandas scikit-learn xgboost torch yfinance

from iching_model import IChingModel

# Define which features belong to each Vital
model = IChingModel(
    tian_label=["market_index_return", "sector_etf_return"],   # Ten / macro
    di_label=["volume", "turnover_rate", "fund_flow"],         # Chi / flow
    ren_label=["net_profit_margin", "debt_ratio", "roe"],      # Jin / fundamentals
    label_window=60,      # rolling window for Huber label construction
    window_size=8,        # Four Signs sliding window (paper default)
    n_top_total=30,       # RF top-N feature selection
    n_top_per_vital=2,    # features per Vital (→ 6 total)
)

model.fit(feature_df, close_series)

# Predict with I Ching explanations
results = model.predict_with_explanation(feature_df)
print(results[["label", "explanation"]].tail())

# Evaluate
metrics = model.evaluate(feature_df, close_series)
print(f"F1: {metrics['f1']:.4f}")
print(metrics["report"])

cd iching
python demo.py

Fetches AAPL data from yfinance (2015–2024) and runs the full pipeline. Falls back to synthetic data if yfinance is unavailable.

cd iching
python benchmark.py              # AAPL via yfinance
python benchmark.py --ticker TSLA
python benchmark.py --synthetic  # force synthetic data

Six comparison algorithms are implemented in baselines.py, matching the paper's Table 2 setup:

All share a unified interface:

from baselines import SVMModel, XGBoostModel, RFModel, KNNModel, GRUModel, LSTMAttModel

model = SVMModel()
model.fit(X_train, y_train)       # y in {-1, 0, 1}
preds = model.predict(X_test)     # returns np.ndarray of {-1, 0, 1}

Chronological 80/20 train/test split · 2515 trading days (2015–2024) · 11 features · macro-averaged metrics.

Test label distribution: 97 Negative / 195 Neutral / 200 Positive

Test label distribution: 36 Negative / 261 Neutral / 195 Positive

All models cluster tightly (23–28% F1) on the index. The Neutral-heavy label distribution (53% of test samples) reflects that the S&P 500 spends most of its time drifting sideways. No model reliably catches Negative (bear) periods — only 36 test samples — because a single-ticker feature set lacks the cross-stock signals needed to separate sharp drawdowns from sideways drift.

Note on differences from the paper: The paper uses 3000 Chinese stocks with full quarterly fundamentals and daily money-flow data. This benchmark uses a single ticker with 11 features, 4 of which are synthetic (fundamentals not available from yfinance). The I Ching model's advantage is strongest in the multi-stock, industry-segmented setting — see Table 3 where industry-specific F1 reaches 76.94% (Power and Energy).

A full audit against the paper's three algorithms identified and fixed three critical bugs.

Paper: del sortedvec[m] removes the oldest element by its tracked index position m in the sorted array.

Was: sortedvec.remove(old_val) — removes the first occurrence by value, silently producing wrong results when duplicate values exist in the window.

Fix: Uses bisect.bisect_left to locate the exact index of the oldest value, then del sortedvec[m] by position. Also switched to bisect.insort for O(log n) insertion instead of re-sorting each step.

Paper: "the distribution of a is counted, and finally the label category is obtained according to the distribution interval of a" — mean ± std from the normal distribution.

Was: np.percentile(slopes, 33.33) and np.percentile(slopes, 66.67) — always forced exactly ⅓ of samples into each class regardless of the actual distribution shape.

Fix: Thresholds are now mean ± n_std × std (default n_std=1.0), producing a Neutral-heavy label distribution (~68% within 1σ) consistent with the paper's normal distribution assumption.

Paper: All 6 permutations of the Three Vitals ordering are evaluated; the best-performing Yao sequence arrangement is kept.

Was: The permutation loop regenerated hexagrams from the same fixed six_features_ order on every iteration — all 6 iterations were identical, so the "best" ordering was never actually searched.

Fix: Each permutation now reorders the 6 features by Vital before generating hexagrams. The best ordering is saved as best_six_features_ and used by predict() and predict_with_explanation().

Liu, W., Chen, S., Huang, G., Lu, L., Li, H., & Sun, G. (2023). Incorporating I Ching Knowledge Into Prediction Task via Data Mining. Journal of Database Management, 34(3). https://doi.org/10.4018/JDM.322097