# 期末報告 ###### tags: `報告` - 報告需包含 data preprocessing、 model architecture、performance analysis - 10 分鐘 - PPT: https://reurl.cc/Gob4OG - PPT(01_13): https://reurl.cc/Vj52RA # Outline 1. Data Pre-processing 2. Model Architecture 3. Performance Analysis 4. Conclusion # Data Pre-processing ### Missing value - 添補 Timestamp 的空缺 - 60 秒一個 Timestamp - 用前後Timestamp填補 - Ex: 1632122520 ~ 16321225640，中間缺 1 個 Timestamp - 添補 Features 的空缺 - 先用前值添補 - 若還有空缺，則用後值補 ### Generate More Features 有些模型有額外生成更多特徵，生成的特徵有 - spread : $High - Low$ - mean_trade : $Volume / Count$ - log_price_change : $\log(Close-Open)$ - upper_Shadow : $High - max(Close,Open)$ - lower_Shadow : $min(Close,Open) - Low$ - high_div_low : $High / Low$ - trade : $Close - Open$ ### Generate Training Data - Model's input/output format: - X: All features except `Target` - Y: two types - Only `Target` feature - All features - Two scalers will be used in different model: - StanderScaler: value`-1`~`1` - MinMaxScaler: value`0`~`1` ### Generate Time Series Data - Time range - 2021.06 ~ - Time series size(step) - 3,10,15 # Model Architecture - Single Model - Many to Many - Many to One - 2 layers LSTM (600) - Dropout : 0.2 - Early stopping - Multi-Model - Many to Many - 2 layers LSTM (8) - Dropout : 0.2 # Performance Analysis | Model | Training Time | Testing Time | Kaggle Score | | -------------------------------- | ------------- | ------------ | ------------ | | Single model </br> (Many to One, 10 steps) | 313.8s | < 9hr | 0.1131 | | Single model </br> (Many to Many, 3 steps)| 337.5s | < 9hr | 0.0495 | | Multi-model </br> (Many to Many, 15 steps)| 255.0s | > 9hr | Timeout | # Conclusion ### Bottleneck - Time Limit - 將預測的筆數(3個月): 3x30x24x60 = 129,600 - 時間限制(9個小時) : 9*60 = 540(分鐘) - (540/129600) * 60 = 0.25秒(預測每筆資料的時間限制) - Single model (P100) 約 0.1 秒 - Multi-model (P100) 約 0.6 秒 -> Timeout - Memory Limit - Training data Timestamp range : last half year - Time series : 3 ~ 15 steps - Features : add 7 columns ### Q & A ### 報告負責 1. 1-6 2. 7-13 3. 14-16 4. QA