工業技術研究院

技術簡介

虛擬理財專員系統，結合市場交易數據與經濟指標，透過深度學習演算法與財務工程理論預測市場投資方向與波動度，並以機器學習中的整體式學習法，建立潛力基金預測模型。透過機器人觀點篩選市場並取得市場下潛力基金後，透過現代投資理論進行資產配置與動態重配置，藉此降低投資組合風險。此外，我們透過深度增強式學習演算法將市場預測結果作為模型輸入，訓練自動化決策模型，針對單一基金定期(不)定額投資建議停扣、贖回、加減碼等投資操作。

Abstract

The techniques of deep learning and financial engineering are integrated for predicting the future market direction and volatilities. Transaction data and economic indicators are used for training these models. Moreover, an ensemble learning algorithm is developed for finding the potential funds in a specific market. With these analytic results, the robo can select the potential markets and obtain the potential funds from these specific markets. Then, it generates an asset allocation with the minimum volatility under a given expected return according to the Markowitz modern portfolio theory, and dynamically reallocates the portfolio while the robo views change. Moreover, we use the deep reinforcement learning algorithms which take the market prediction results as inputs to make the investment decisions for investing a single mutual fund in the way of variant dollar-cost averaging.

技術規格

1. 本技術支援至少33個以上全球股債市場之潛力基金分析、投資方向分析與波動度分析。市場投資方向分析包含股票市場未來一個月、三個月、六個月投資方向分析，與債券市場未來三個月、六個月、一年之投資方向分析，市場投資方向準確度達7成以上；波動度分析包含股票市場未來一個月波動度相對當前波動度變化程度分析。 2. 本技術支援市場潛力基金分析，潛力基金平均報酬贏過市場平均基金報酬之機率為0.67，未來一個月之報酬率改善率41.59％。 3. 本技術支援定期(不)定額AI模型自動化投資操作，可由智慧決策是否贖回、續扣、加減碼等投資操作。

Technical Specification

1. By using structured data, e.g. market transactional data and economic indicators, and unstructured data, e.g. reports and monetary policies, we train and apply multi-step-ahead learning models to predict the future short term and long term directions of the stock and bond indices, resulting in an average accuracy exceeding 70％, say the directions of the next 1, 3, 6 months for stock indices and those of the next 3, 6, 12 months for bond indices. In addition, GARCH-MIDAS and the Monte Carlo simulation are used to evaluate the future volatilities of stocks, i.e., the volatility change of the next month. 2. We support the potential fund analytics. The returns of the selected potential funds of the markets outperform the average returns of the funds in a chance of 67％. The improvement rate of the next month return is 41.59％. 3. Many investors choose to invest in specific mutual funds using the dollar-cost averaging strategy, putting small amounts of money every month to meet the amount limitation policy of a mutual fund and investing in several funds simultaneously. We develope an investment agent for this purpose, using deep reinforcement learning algorithms and taking market direction and volatility predict into account, to make purchase decisions.

技術特色

結合市場交易數據與經濟指標，透過深度學習與財務工程理論預測50個股債市場之市場投資方向與波動度，並以機器學習中的整體式學習法建立潛力基金預測模型。透過機器人觀點篩選市場並取得市場下潛力基金後，透過現代投資理論進行資產配置與動態重配置，藉此降低投資風險。此外，我們透過深度增強式學習演算法將市場預測結果作為模型輸入，訓練自動化決策模型，針對單一基金定期(不)定額投資建議停扣、贖回、加減碼等投資操作。

應用範圍

智慧理財機器人可應用於數位金融服務，對金融企業內部提供市場與投資標的分析結果，協助理專進行客戶經營。

接受技術者具備基礎建議（設備）

軟體：Spark 2.2.1、Keras 2.1.2、TensorFlow 1.2.1、Keras-rl 0.4.2以上版本。 硬體：叢集環境，至少提供32 cores與256GB記憶體作為運算所需之計算資源。NVIDIA Tesla V100 GPU。

接受技術者具備基礎建議（專業）

具金融背景知識，具資料分析與深度學習演算法基本概念，以及python開發能力。

聯絡資訊