• Àüü
  • ÀüÀÚ/Àü±â
  • Åë½Å
  • ÄÄÇ»ÅÍ
´Ý±â

»çÀÌÆ®¸Ê

Loading..

Please wait....

±¹³» ÇÐȸÁö

Ȩ Ȩ > ¿¬±¸¹®Çå > ±¹³» ÇÐȸÁö > µ¥ÀÌÅͺ£À̽º ¿¬±¸È¸Áö(SIGDB)

µ¥ÀÌÅͺ£À̽º ¿¬±¸È¸Áö(SIGDB)

Current Result Document : 20 / 20

ÇѱÛÁ¦¸ñ(Korean Title) ½ÉÃþ°­È­ÇнÀ ±â¹ÝÀÇ ½Ç¿ëÀûÀÎ Æä¾î Æ®·¹À̵ù ¹æ¹ý
¿µ¹®Á¦¸ñ(English Title) A Practical Pairs-Trading Method Using Deep Reinforcement Learning
ÀúÀÚ(Author) ±è»óÈ£   ¹Ú´ö¿µ   À̱âÈÆ   Sang-Ho Kim   Deog-Yeong Park   Ki-Hoon Lee  
¿ø¹®¼ö·Ïó(Citation) VOL 37 NO. 02 PP. 0065 ~ 0080 (2021. 08)
Çѱ۳»¿ë
(Korean Abstract)
Æä¾î Æ®·¹À̵ù(pairs trading)Àº À¯»çÇÏ°Ô ¿òÁ÷ÀÌ´Â µÎ ÁÖ½Ä Á¾¸ñÀÇ ´Ü±âÀû °¡°Ý Â÷ÀÌ(½ºÇÁ·¹µå)¿Í Æò±Õ ȸ±Í¼ºÀ» ÀÌ¿ëÇÏ´Â ÅõÀÚÀü·«ÀÌ´Ù. ÃÖ±Ù ½ÉÃþ°­È­ÇнÀÀ» ÀÌ¿ëÇÑ Æä¾î Æ®·¹À̵ù ¿¬±¸°¡ ÀÌ·ç¾îÁö°í ÀÖÀ¸³ª, °ø¸Åµµ°¡ °¡´ÉÇÑ ½ÃÀåÀ» °¡Á¤ÇÏ°í ÀÖ°í ÀâÀ½ °¨¼Ò ¹× Ư¡ ÃßÃâ °úÁ¤ÀÌ ´Ü¼øÇÏ¿© ÀϹÝÈ­°¡ ¾î·Æ´Ù. º» ³í¹®¿¡¼­´Â ±âÁ¸ ¹æ¹ýÀÇ ¹®Á¦Á¡À» ÇØ°áÇÑ ½ÉÃþ°­È­ÇнÀ ±â¹ÝÀÇ ½Ç¿ëÀûÀÎ Æä¾î Æ®·¹À̵ù ¹æ¹ýÀÎ P-Trader¸¦ Á¦¾ÈÇÑ´Ù. Á¦¾ÈÇÑ ¹æ¹ýÀº ½ºÇÁ·¹µå¿¡ ´ëÇÑ ±â¼úÀû ÁöÇ¥¸¦ °è»êÇÏ°í ½ºÇÁ·¹µå ĵµé½ºÆ½(candlestick)À» ±ºÁýÈ­ÇÏ´Â µ¥ÀÌÅÍ Àüó¸® ÀÛ¾÷À» ¼öÇàÇÑ´Ù. Àüó¸®µÈ Ư¡¿¡ °¡ÁßÄ¡¸¦ ºÎ¿©ÇÏ°í Gated Recurrent Unit (GRU)°ú ÁÖÀÇÁýÁß ¹æ¹ý(attention mechanism)À» Àû¿ëÇÏ¿© ½Ã°£Àû Ư¼ºÀ» ¹Ý¿µÇÑ »óÅ º¤Å͸¦ »ý¼ºÇÑ´Ù. »óÅ º¤Å͸¦ ±â¹ÝÀ¸·Î Double Deep Q-Network (DDQN) ±â¹ÝÀÇ °­È­ÇнÀÀ» Àû¿ëÇÑ´Ù. °­È­ÇнÀ °úÁ¤¿¡¼­ ´ÙÀ½ ½ÃÁ¡ÀÇ ½ºÇÁ·¹µå °ªÀ» ¿¹ÃøÇÏ´Â ¸ðµ¨À» ÇÔ²² ÇнÀ½ÃÄÑ °­È­ÇнÀ ¸ðµ¨¿¡ Á¤È®ÇÑ »óÅ Á¤º¸¸¦ Á¦°øÇÑ´Ù. KOSPI Á¾¸ñ¿¡¼­ 7°³ÀÇ Æä¾î¸¦ ¼±Á¤ÇÏ¿© ½ÇÇèÇÑ °á°ú, Á¦¾ÈÇÑ ¹æ¹ýÀÌ ¸ðµç Æä¾î¿¡ ´ëÇØ ±âÁ¸ ÃֽŠ¹æ¹ýµéº¸´Ù ´õ ³ôÀº ¼öÀÍ·üÀ» º¸¿´´Ù.
¿µ¹®³»¿ë
(English Abstract)
Pairs trading is an investment strategy that uses the short-term price difference (spread) between two co-moving stocks and mean reversion. Recently, there have been studies on pairs trading using deep reinforcement learning. However, the studies assume a market allowing short-selling, and their process of noise reduction and feature extraction is not sophisticated, resulting in low generalization ability. In this paper, we propose a practical pairs-trading method called P-Trader based on deep reinforcement learning that overcomes the limitation of existing studies. We preprocess data by computing technical indicators for the spread and clustering the candlesticks of the spread. We give weights on the preprocessed features and apply Gated Recurrent Unit (GRU) and an attention mechanism to generate state vectors with temporal characteristics. Based on the state vectors, we apply reinforcement learning based on Double Deep Q-Network (DDQN). During reinforcement learning process, we co-train a model that predicts the next spread value to provide accurate state information to the reinforcement learning model. The experimental results using 7 stock pairs in KOSPI show that the proposed method accomplishes better profit for all the pairs compared with the state-of-the-art methods.
Å°¿öµå(Keyword) ¾Ë°í¸®Áò Æ®·¹À̵ù   Æä¾î Æ®·¹À̵ù   ½ÉÃþ°­È­ÇнÀ   Double Deep Q-Network   Algorithmic Trading   Pairs Trading   Deep Reinforcement Learning   Double Deep Q-Network  
ÆÄÀÏ÷ºÎ PDF ´Ù¿î·Îµå