Portfolio construction turns investment signals into holdings, balancing expected return against risk, diversification, and trading costs. In a daily portfolio-level system, an RL agent combines a news signal with price features to choose what to hold. Execution is a separate stage; this system excludes it and assumes rebalances fill at the closing price. section 12.3
An allocation decision
The agent chooses among eight allocation templates, such as equal-weight, defensive, or growth/tech. Its daily reward accounts for portfolio return, transaction costs, and penalties for turnover, drawdown, and volatility. section 12.3
Table 12.1 The portfolio-management decision as a Markov decision process
|
Element
|
In This System
|
|
State
|
Market price features, a
news-signal block, portfolio holdings, and execution/risk context (63 numbers
in total)
|
|
Action
|
One of eight named allocation
templates: cash-heavy, defensive, equal-weight, growth/tech, energy,
financials, broad risk-taking, or risk-avoiding
|
|
Reward
|
Daily portfolio return minus
transaction costs and penalties for turnover, drawdown, and volatility
|
|
Episode
|
A window of 30 consecutive trading
days
|
|
Transition
|
Advance one day, realize the next
day’s return, and update portfolio holdings, value, and drawdown
|
See how market and portfolio information becomes an RL allocation decision, including its state, action, reward, episode, and transition. section 12.3
