Inventory Capital Optimizer

# Custom Metric: SPEC (Stock-keeping-oriented Prediction Error Costs)
def calculate_spec_grouped(y_true, y_pred, product_ids):
    """
    Calculates cumulative cost of errors per SKU to prevent cross-product pollution.
    Penalizes stockouts (0.75) vs overstocks (0.25).
    """
    eval_df = pd.DataFrame({'PID': product_ids, 'Actual': y_true, 'Pred': y_pred})
    sku_costs = []

    for _, group in eval_df.groupby('PID'):
        act, pre = group['Actual'].values, group['Pred'].values
        n = len(act)
        if n == 0: continue
        
        cum_act, cum_pre = np.cumsum(act), np.cumsum(pre)
        cost = 0.0
        
        for t in range(n):
            # Calculate unfulfilled demand and excess inventory
            unfulfilled = np.maximum(0, cum_act[:t+1] - cum_pre[t])
            excess = np.maximum(0, cum_pre[:t+1] - cum_act[t])
            
            # Weighted penalty: Stockout risk is 3x more costly than trapped capital
            penalties = np.maximum(unfulfilled * 0.75, excess * 0.25)
            cost += np.sum(penalties * (t - np.arange(t+1) + 1))
            
        sku_costs.append(cost / n)
        
    return np.mean(sku_costs)

# Example Usage during model tournament
spec_score = calculate_spec_grouped(y_test, y_pred, test_ids)
print(f"SPEC Optimizer Score: {spec_score:.4f}")