定義
過度擬合
過度擬合係將策略調到貼合歷史噪音而唔係真實訊號。本文說明佢嘅警示訊號，以及點解佢係優化嘅預設結局。
更新於
過度擬合係指將一個策略反覆調參，直到佢描述嘅係歷史樣本中嘅噪音，而唔係市場中任何可重複嘅結構。佢嘅特徵係：一旦換到「參數唔係由入面揀出嚟」嘅數據上，表現就崩。呢個唔係一種罕見嘅失敗模式，而係無約束優化嘅預設結局：只要參數夠多、嘗試次數夠多，任何一份數據都可以被完美擬合。
- 又稱
- 曲線擬合 · 數據窺探
計算方式
Symptoms of an overfitted strategy:
· Performance collapses on data the parameters were not chosen from
· Small parameter changes produce large performance changes
· More rules and more conditions than the sample can support
· An equity curve that is smooth in-sample and jagged out-of-sample
你跑過嘅每一次回測都係有代價嘅
測試一百個變體然後留低最好嗰個，即使呢一百個裡面冇一個真係有優勢，你都一定會得到一份靚嘅樣本內結果——因為勝出者係被噪音揀出嚟嘅。呢個就係點解「試過幾多組配置」必須被記錄並折價處理，亦係點解一個捱得住一次誠實檢驗嘅策略，比一個由一千次裡面勝出嘅策略更有說服力。
穩健嘅策略睇落好悶
當參數偏離最優點時，表現係平滑衰減而唔係斷崖式崩塌嘅策略，更有可能真係描述咗某種現實存在嘅嘢。如果回看 5 個週期賺錢、回看 6 個週期就唔賺，咁呢個策略搵到嘅係樣本嘅一個偶然產物，而唔係市場嘅一個屬性。
常見誤讀
樣本外測試經常被當作解藥。但佢只係一次檢驗，而且喺你用佢嘅結果去修改策略嗰一刻就失效：一旦你睇過留出集並據此改咗啲嘢，呢份留出集就變成咗樣本內數據，嗰份誠實嘅證據亦冇咗。
親手計一計
參考資料
- Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance — Notices of the American Mathematical Society (Bailey, Borwein, López de Prado, Zhu)
- … and the Cross-Section of Expected Returns — Campbell R. Harvey, Yan Liu, Heqing Zhu — Duke University (published in the Review of Financial Studies)
- Data-Snooping Biases in Tests of Financial Asset Pricing Models — National Bureau of Economic Research (Andrew W. Lo, A. Craig MacKinlay)
本術語表用於教學。此處內容不構成投資建議，頁面所述任何指標都無法預測未來結果。