DifferentialSharpeReward
Models & Types · AiDotNet.Finance.Trading.Rewards
Differential Sharpe Ratio reward (Moody & Saffell, 1998) — the canonical risk-adjusted objective for RL
trading. It rewards the MARGINAL contribution of each step’s return to the online Sharpe ratio, so the agent
is pushed toward smooth, consistent gains rather than volatile ones (a big gain that spikes variance can
score LOWER than a smaller steady gain). Maintains exponentially-weighted first (A) and second (B) moments of
the return with decay eta; the differential is
D_t = (B·ΔA − 0.5·A·ΔB) / (B − A²)^{3/2}, which is the derivative of the Sharpe ratio w.r.t. the EWMA
weight — a per-step, online, backprop-free shaping signal. An optional turnover penalty is subtracted.
Implements: IPortfolioReward
Constructors
| Constructor | Summary |
|---|---|
DifferentialSharpeReward(double eta, double turnoverPenalty) |
Methods
| Method | Summary |
|---|---|
double Reward(PortfolioRewardContext context) |
|
void Reset() |