在2020 年初,由於Covid-19 的大流行,全球股票市場面臨了災難般地崩跌。這反映了即使處在看漲的「牛市」中,股票價格也隨時都可能崩盤。 在此研究中,我們希望透過避免未來的「價格修正」以獲取超額報酬,藉此改進「買入並持有」的投資策略。我們設計了「價格修正模型」,通過調整此模型的門檻值之組合,此標記演算法便可以精確地針對不同資產標記其「價格修正」。 接著,我們提出了一個2D GADF-CNN 模型以學習「價格修正」之間的共通規律。股票時間序列會先轉換為GAF 矩陣,再輸入到此模型中。在所有模型中,給定NASDAQ 指數最大的ETF-QQQ 之資料,CNN 模型在統計指標和回測報酬率上都表現最好。 最後,為了進一步測試我們模型的強健性,我們將其套用在TSM 和TSLA的資料上,分別為和QQQ 相似以及不相似的股票。2021 年1 至3 月的回測結果顯示,無論是直接對相似資料集進行遷移學習,亦或是針對相似與不相似的資料集微調「價格修正模型」的門檻值,我們的模型皆能習得有用的規律進而避開未來的下跌趨勢,最終得到超越「買入並持有」的投資報酬率。
Due to the Covid-19 pandemic, stock market all around the world face a disastrous drop down at the beginning of 2020. This indicates that even in the bullish market, stock price may collapse in every moment. In this study, we aim to improve “Buy and Hold” Strategy, through avoiding potential “Price Corrections” to gain excess return. We proposed “Price Correction Model”, which is a labeling algorithm could precisely characterize “Price Corrections” of each asset, by different thresholds given. Next, a 2D GADF-CNN model is proposed to learn generic patterns between corrections. Stock time series will first encode as GAF matrix, then input to the model. With the biggest ETF of NASDAQ index – QQQ as input, the CNN model performs best on both statistic indicators and back-testing return between all models we compare. Finally, in order to test the robustness of our model, we applied our model to TSM and TSLA, which is a similar case and a disparity case of QQQ, respectively. The back-testing during Jan. 2021 to Mar. 2021, shows that by directly transfer learning on similar dataset, or fine-tuned thresholds of “Price Correction Model” on both similar and disparity case, our model learns useful patterns to avoid future draw downs, and gain excess return beyond “Buy and Hold” Strategy.