CompetiLab | Quant Trading Technical Paper

Abstract

研究問題

在交易成本、滑價、walk-forward、holdout 與回撤限制都納入後，是否能建立一組可重跑、可審查、跨商品的自動交易策略，並在明確 baseline 下取得風險調整後的優勢？

核心發現

最穩定的結果不是單一模型勝出，而是「資料協議 + 訊號族群 + 風控 overlay + 嚴格 promotion gate」的組合。期貨策略相對 live hurdle 有顯著超額報酬；台股 Top200 alpha 相對 0050 有正的 total-return 與 OOS CAGR gap。

Best futures 5y return7,712.08%

Top200 vs 0050 total gap+331.19%

Risk-sweep MaxDD reduction-4.10pp

CZSC rigor gate13 / 13

1. Data Acquisition & Target Instruments

資料先定義成 protocol，再允許策略讀取。

每個資料集先固定 schema、時間區間、可交易時間點與成本模型；策略只能讀取符合 contract 的資料。

01

Market Bars

台灣指數期貨與小型期貨的 1m / 5m OHLCV bars，按台北時區校準，切分 regular / night session，保留日內停損與隔日延遲填單需要的時間欄位。

02

Equity Universe

台股 Top200 historical universe 與 0050 benchmark，使用 walk-forward universe selection，避免把未來成分股或未來 benchmark 狀態洩漏到訓練期。

03

Exogenous Features

ADR、macro、volatility、trend、session regime 等外部特徵全部以可交易時間點前可得的 lagged values 進入模型，禁止 forward fill 到未來事件。

04

Execution Model

回測先扣 commission、futures tax、slippage、stop slippage，並將 next-bar fill、延遲填單、日損限制與合約上限納入同一套 replay。

Target Instrument	Research Objective	Primary Baseline	Core Metrics
TX / MXF intraday futures	Find high-conviction event and regime strategies under realistic costs.	Live hurdle: 4,347% 5y return, Sharpe 1.46, MaxDD 10.04%.	5y return, Sharpe, MaxDD, 2025+ holdout Sharpe.
Taiwan Top200 equities	Build benchmark-aware stock selection against broad-market exposure.	0050 ETF benchmark.	Total return, CAGR, Sharpe, MaxDD, OOS CAGR gap.
CZSC low-frequency MXF cells	Convert chart-structure signals into lower-turnover futures sleeves.	Buy-and-hold / incumbent futures sleeve.	Train / holdout points, rigor gates, monthly stability.

2. Methodology

策略不是單一模型，而是可審查的研究鏈。

每個 strategy family 都必須交代訊號來源、selection rule、risk overlay、驗證 gate 與失敗條件。

Figure 2. Method stack: data contract, signal family, selector, risk overlay, validation gate.

Signal Families

Markov state, price / macro regime, CZSC structure, CCI daily, trend ensemble, time-exit repricing, and forecast overlay.

Selection Layer

Walk-forward folds, edge buckets, Markov prior, Bayesian shrinkage, LCB ranking, and benchmark-aware stock selection.

Risk Layer

Contract cap, Kelly fraction, daily loss limit, emergency DD, monthly loss stop, overlay haircut, and drawdown rescue.

Strategy Family	Design Rationale	Validation Role	Public Verdict
MHF 5m price-macro	Combine state transition and macro-sensitive price structure to isolate high-conviction futures events.	Main futures alpha candidate under strict hurdle.	Verified pass
Overlay / haircut ensemble	Reduce exposure around known drawdown clusters without changing the base signal.	Risk-control proof and ablation target.	Verified pass
Top200 trend ensemble	Select stock sleeves only when benchmark-aware trend evidence survives walk-forward selection.	Equity alpha vs 0050.	Verified pass, deployment gated
Kronos forecast overlay	Fine-tune a time-series foundation model per walk-forward fold, then use forecasts as position scaling rather than standalone orders.	Implementation example below; research candidate due to true-DD caveat.	Candidate
Regime / DD gate sweep	Search only risk-control coordinates after the signal is frozen.	Controls overfitting by limiting the search space.	Verified pass

3. Implementation Example: Kronos Fine-Tuning

Kronos 的用法不是直接喊多空，而是產生下一根 bar 的 forecast overlay。

流程重點是 fold 隔離、固定輸入 schema、低學習率 fine-tune、forecast 延遲生效，以及把模型輸出放進既有風控 replay。

Input

5m OHLCV contract

輸入欄位固定為 timestamp、open、high、low、close、volume、amount。時間以台北時區校準，night session 與 regular session 不混用未來資訊。

Split

Half-year OOS folds

每個測試 fold 約半年；訓練資料只到 cutoff 為止。2021-2026 combined audit 覆蓋 11 個 folds，產生 298,253 筆 forecast rows。

Target

One-bar horizon

lookback_window = 512、predict_window = 1。模型只預測下一根 5m bar 的相對方向分數，不直接下單。

Replay

Overlay, not signal replacement

Kronos 分數只調整既有部位大小；所有交易仍經過 contract cap、日損限制、emergency drawdown 與成本滑價 replay。

Stage	Concrete Setting	Why It Matters
Pretrained base	Kronos tokenizer base + Kronos small predictor.	把 foundation model 當成可微調的 sequence prior，而不是重新訓練大型模型。
Windowing	512-bar context, 1-bar prediction, value clip = 5.0.	限制每次樣本只看固定長度歷史，並降低極端價格變動對 tokenizer 的影響。
Train / validation	90% train, 10% validation inside each fold, no test rows inside fine-tune data.	測試區間只能用於 forecast replay，不能回流到 fine-tune 或 model selection。
Optimization	Tokenizer 1 epoch, predictor 1 epoch, batch size 8, learning rate 1e-5.	用低學習率做 domain adaptation，避免在單一 fold 上過度改寫 pretrained weights。
Adam settings	beta1 = 0.9, beta2 = 0.95, weight decay = 0.1, seed = 17.	固定 optimizer 與 random seed，讓不同 folds 的差異主要來自資料而不是訓練漂移。
Audit rule	trained_through_ts must be no later than fold cutoff; effective_from_ts must be after decision_ts.	這是防止 look-ahead 的硬檢查；不通過就不進入結果表。

Public pseudo-config Kronos fold runner

input_schema = ["timestamp", "open", "high", "low", "close", "volume", "amount"]

for fold in walk_forward_folds:
    train = bars.where(timestamp <= fold.cutoff)
    test = bars.where(fold.start <= timestamp < fold.end)

    model = KronosSmall.from_pretrained()
    model.fit(
        train,
        lookback_window=512,
        predict_window=1,
        train_ratio=0.90,
        val_ratio=0.10,
        batch_size=8,
        epochs={"tokenizer": 1, "predictor": 1},
        learning_rate=1e-5,
        weight_decay=0.1,
        seed=17,
    )

    forecasts = model.predict(test)
    assert forecasts.trained_through_ts.max() <= fold.cutoff
    assert all(forecasts.effective_from_ts > forecasts.decision_ts)

Forecast table

Replay only accepts timestamped forecasts

每筆輸出包含 decision_ts、effective_from_ts、trained_through_ts、score、horizon_bars、confidence、model_id。缺任何一個時間欄位就不能進 replay。

Position scaling

Forecast aligns: amplify; contradicts: no amplify

實驗版使用 inverted daily forecast feature；同向時把 target exposure 放大到 overlay scale，反向、中性或缺資料時回到原始 exposure，最後再套 max_abs_position cap。

Observed replay

Candidate result, not deployment claim

Daily overlay research replay: 7,073.48% return、15.24% MaxDD、998 forecast days。Event replay: 5,059.10% return、monthly Sharpe 1.526，但 true-DD caveat 仍需處理。

4. Baseline Definition

每個結果都必須先指定比較對象。

期貨策略比較 live hurdle；股票策略比較 0050。若沒有 baseline，就不宣稱「贏」。

Figure 3a. Futures strategy returns relative to the live hurdle.

Figure 3b. Top200 strict ensemble relative to 0050 benchmark.

Futures baseline 4,347% return / 10.04% MaxDD

Any futures claim must survive return, Sharpe, drawdown, and holdout gates after cost and slippage.

Equity baseline 0050 ETF

Any Top200 claim must show benchmark-aware excess return and OOS strength, not only absolute bull-market beta.

5. Results

結果以表格呈現：策略、baseline、metric、verdict。

表格只接受已指定 baseline、成本模型、回撤口徑與驗證狀態的結果；candidate 不視為 production claim。

Experiment	Baseline	Result	Risk	Conclusion
MHF 5m fixed baseline	Live hurdle 4,347%	7,712.08% 5y return, Sharpe 2.993	MaxDD 8.085%	Primary verified futures result.
MHF Markov + Deep	Same hurdle	5,194.43% 5y return, Sharpe 2.732	MaxDD 9.895%	Model selection passes strict hurdle.
Overlay haircut ensemble	Base signal and live hurdle	5,782.78% 5y return, Sharpe 2.098	MaxDD 9.764%	Risk overlay survives ablation and neighborhood checks.
Parity DD-gate sweep	Live config MaxDD 10.04%	4,452.10% return, daily Sharpe 1.592	MaxDD 5.94%	Valid risk-control improvement after signal is frozen.
Top200 strict ensemble	0050 total return 515.14%	846.33% total return, CAGR 30.01%	MaxDD 28.65% vs 0050 33.83%	Verified equity alpha with positive OOS CAGR gap.
CZSC 5-cell MXF portfolio	Single-cell / low-frequency sleeve	12,352 holdout points, Sharpe 6.8	13 / 13 rigor pass	Validated low-frequency structure portfolio.
Kronos TX overlay	Live-risk replay	5,059.10% event replay return	Reset DD 6.11%, true DD caveat	Fine-tuned overlay candidate; not broker-ready.

6. Validation & Failure Rules

驗證重點是防止漂亮回測誤導。

任何策略若只有 return 而缺少成本、holdout、回撤、ablation 或 execution identity，會被降級為 candidate 或 rejected path。

WF

Walk-forward split

固定 train / test windows，讓 selection rule 在每個 fold 中只看過去資料；Kronos audit 要求 fold leakage violations = 0。

OOS

Holdout period

2025+ 或 predefined out-of-sample window 必須維持 Sharpe、return、drawdown 的基本品質。

COST

Cost-first replay

Commission、tax、slippage、stop slippage 先進入 replay，再談任何績效宣稱。

ABL

Ablation and local scan

Overlay 類策略需要移除單一元件後仍能說明其貢獻，並檢查不是單點參數幸運。

DD

Drawdown honesty

Reset drawdown 與 true drawdown 分開報告；若 true DD 過大，不升級為 production claim。

FAIL

Rejected paths are retained

未過 strict gate 的 15m / 5m walk-forward、independent replacement、direct fusion 都只保留為 guardrail。

7. Conclusion

結論：可交易研究的核心是「可重跑的實驗設計」，不是單一高報酬數字。

本研究支持一個實務結論：跨商品自動交易要先建立資料協議、baseline、風控與驗證 gate，再讓策略族群競爭。MHF、overlay ensemble、Top200 strict ensemble 與 CZSC portfolio 顯示了可驗證的優勢；Kronos fine-tuning 則提供一個可重跑的 forecast overlay 範例，但仍需更嚴格的 true-DD 與 deployment parity 檢查。