Volatility · Advanced · Daily
Long Straddle Options
How it works, and when it doesn't.
The Q-table converges to cash during high-volatility unclear regimes and long/short during confident directional states — a behavior that emerges purely from reward maximization without explicit programming.
Why it works
A tabular Q-learning agent learns an optimal trading policy for 3 actions (Long, Short, Cash) from market states. Unlike supervised models, the RL agent learns directly from trading profits — it discovers that staying cash during high-uncertainty states and riding momentum during clear trends maximizes reward.
The signal logic
- State: discretized (trend strength, vol level, momentum) — 27 states
- Actions: Long (1), Cash (0), Short (-1)
- Q-table: Q(s,a) += lr * (reward + γ*max_Q(s') - Q(s,a))
- Policy: epsilon-greedy with decay (explore then exploit)
- Signal from greedy policy (highest Q-value action for current state)
When it fails
Tabular Q-learning cannot generalize to unseen states. The coarse state space misses subtleties captured by continuous models.
- Best market
- Markets with persistent momentum where the agent learns to ride trends
- Worst market
- Highly non-stationary markets where the learned policy quickly becomes stale
Backtest
This is a research tool rather than a directional strategy; it has no live signal of its own.
Long Straddle Options, live.
Re-scored after every close.
The platform runs this and 32 other strategies against the live tape each day — with the evidence behind every number.