# 美股籌碼評分系統 — 12 個月校準報告

**生成時間**:2026-05-15T20:28:52  
**Universe**:503 檔(S&P 500)  
**樣本**:4017 觀察值(月度)  
**訓練 / 驗證**:3012 / 1005  (cutoff = 2025-09-30)  
**校準目標**:6 個月 forward return

## 摘要

建議**保留現有權重**:Grid search 找不到顯著更好的(test IC +0.0925 vs +0.0925)


## 訊號衰減節奏(baseline IC,test set)

| 持有期 | Train IC | Test IC | 解讀 |
|--------|----------|---------|------|
| 1M | -0.0354 | +0.0332 | 短線(雜訊高,通常無預測力) |
| 3M | +0.0249 | +0.0959 | 中期 |
| 6M | +0.0693 | +0.0925 | 長線(本系統設計目標) |

> **解讀**:IC > 0.05 通常被認為有用。**長期投資者最關心 6 個月 IC**。


## 子訊號獨立 IC(全樣本)

| 訊號 | IC | 含意 |
|------|-----|------|
| Insider 訊號(關閉 SI/borrow) | +0.0608 | |
| SI 訊號(關閉 insider) | +0.0559 | |

> 子訊號 IC 顯示**哪一塊預測力最強**。此結果可決定是否要加重某一邊的權重。


## 籌碼分 quintile vs 報酬(test set)

**Baseline 權重 — Test set quintiles**

| Bucket | n | Score range | Mean Return | Median Return |
|--------|---|-------------|-------------|---------------|
| Q1 | 201 | [25, 47] | **+1.52%** | -3.34% |
| Q2 | 201 | [47, 47] | **+5.07%** | +4.12% |
| Q3 | 201 | [47, 47] | **+5.87%** | +5.45% |
| Q4 | 201 | [47, 47] | **+9.49%** | +3.56% |
| Q5 | 201 | [47, 72] | **+7.60%** | -0.23% |


```
        Mean 6m return
  Q1  |                              ████                           +1.52%  (n=201)
  Q2  |                              ██████████████                 +5.07%  (n=201)
  Q3  |                              ████████████████               +5.87%  (n=201)
  Q4  |                              ███████████████████████████    +9.49%  (n=201)
  Q5  |                              █████████████████████          +7.60%  (n=201)
        ------------------------------|------------------------------
                       -10%                       0                       +10%
```

> 理想呈現:Q1 < Q2 < Q3 < Q4 < Q5(單調遞增)。**Q5 - Q1 spread 是最常用的綜合指標**。


## Grid Search 最佳權重

- Train IC (ret_6m):**+0.0720**
- Test IC (ret_6m):**+0.0925**

**權重改動:**

| Weight | Baseline | Best | Δ |
|--------|----------|------|---|
| `insider_buy_cluster_2plus` | 25 | **15** | -10 |
| `insider_buy_cluster_3plus_with_ceo` | 40 | 40 | +0 |
| `insider_sell_cluster_3plus` | -20 | -20 | +0 |
| `regime_squeeze_setup` | 1.3 | **1.1** | -0.20 |
| `si_decrease_score` | 15 | **20** | +5 |

> 其他未列出的權重(borrow fee 相關、insider 細節參數、regime modifier 等)**保持 baseline 不變**,因為它們不在 grid search 範圍內,或在回測階段(無歷史 SI%、無 borrow fee)無法校準。


**Best-weights test set quintiles:**

| Bucket | n | Score range | Mean Return | Median Return |
|--------|---|-------------|-------------|---------------|
| Q1 | 201 | [25, 47] | **+1.38%** | -4.04% |
| Q2 | 201 | [47, 47] | **+5.33%** | +4.22% |
| Q3 | 201 | [47, 47] | **+5.76%** | +5.03% |
| Q4 | 201 | [47, 47] | **+9.49%** | +3.56% |
| Q5 | 201 | [47, 72] | **+7.60%** | -0.23% |



## Grid Search Top 10

| Rank | Train IC | Weights |
|------|----------|---------|
| 1 | +0.0720 | `{"insider_buy_cluster_3plus_with_ceo": 40, "insider_buy_cluster_2plus": 15, "insider_sell_cluster_3plus": -20, "si_decrease_score": 20, "regime_squeeze_setup": 1.1}` |
| 2 | +0.0720 | `{"insider_buy_cluster_3plus_with_ceo": 40, "insider_buy_cluster_2plus": 15, "insider_sell_cluster_3plus": -20, "si_decrease_score": 20, "regime_squeeze_setup": 1.3}` |
| 3 | +0.0720 | `{"insider_buy_cluster_3plus_with_ceo": 40, "insider_buy_cluster_2plus": 15, "insider_sell_cluster_3plus": -20, "si_decrease_score": 20, "regime_squeeze_setup": 1.5}` |
| 4 | +0.0720 | `{"insider_buy_cluster_3plus_with_ceo": 50, "insider_buy_cluster_2plus": 15, "insider_sell_cluster_3plus": -20, "si_decrease_score": 20, "regime_squeeze_setup": 1.1}` |
| 5 | +0.0720 | `{"insider_buy_cluster_3plus_with_ceo": 50, "insider_buy_cluster_2plus": 15, "insider_sell_cluster_3plus": -20, "si_decrease_score": 20, "regime_squeeze_setup": 1.3}` |
| 6 | +0.0720 | `{"insider_buy_cluster_3plus_with_ceo": 50, "insider_buy_cluster_2plus": 15, "insider_sell_cluster_3plus": -20, "si_decrease_score": 20, "regime_squeeze_setup": 1.5}` |
| 7 | +0.0720 | `{"insider_buy_cluster_3plus_with_ceo": 30, "insider_buy_cluster_2plus": 15, "insider_sell_cluster_3plus": -20, "si_decrease_score": 20, "regime_squeeze_setup": 1.1}` |
| 8 | +0.0720 | `{"insider_buy_cluster_3plus_with_ceo": 30, "insider_buy_cluster_2plus": 15, "insider_sell_cluster_3plus": -20, "si_decrease_score": 20, "regime_squeeze_setup": 1.3}` |
| 9 | +0.0720 | `{"insider_buy_cluster_3plus_with_ceo": 30, "insider_buy_cluster_2plus": 15, "insider_sell_cluster_3plus": -20, "si_decrease_score": 20, "regime_squeeze_setup": 1.5}` |
| 10 | +0.0720 | `{"insider_buy_cluster_3plus_with_ceo": 40, "insider_buy_cluster_2plus": 15, "insider_sell_cluster_3plus": -10, "si_decrease_score": 20, "regime_squeeze_setup": 1.1}` |


## ⚠️ 已知限制

1. **NYSE 上市股無 SI 資料** — 本次校準 SI 訊號只覆蓋 NASDAQ-listed 約一半 S&P 500。Insider 訊號全覆蓋。
2. **市值未做歷史回填** — 規模歸一化(insider 買金額 / 市值)用當前市值近似。
3. **借券費(borrow fee)無歷史資料** — 校準階段不採計;production 上線後累積真實資料,6 個月後可二次校準。
4. **Form 4 涵蓋公開市場交易** — 內部人贈與、選擇權執行不歸入訊號(tx_code != P/S)。
5. **樣本期內可能含市場 regime 變化** — 走勢市 vs 修正市的籌碼訊號意義不同。


## 下一步建議

- [ ] 套用建議權重到 `chips_score.py`(若 verdict 顯示有改進)
- [ ] 累積 6 個月 IBKR 借券費歷史後做二次校準
- [ ] 加入 13F 機構持倉變動為第三因子(已有 `flow_collect_13f.py` 基礎)
- [ ] 若 Q5/Q1 spread 顯著,考慮做籌碼分 long-short 模擬績效