Short-Term Trading as a Sequential Decision Problem
1 Sequential decision problems in formal terms
Warren B. Powell's framework characterises every sequential decision problem with five elements:
- State (St) – information required before acting;
- Decision (xt) – the action rule (Xπ(St));
- Exogenous information (Wt+1) – uncertain data that arrives after the decision;
- Transition model (St+1=SM(St,xt,Wt+1)) – system evolution;
- Objective – expected cumulative contribution (E[∑t=0TC(St,xt)]).
This abstraction permits rigorous analysis without over-committing to any single solution technique.
2 How short-term power trading fits the template
| Element | Concrete instance in I-SEM asset-backed trading (wind, solar, BESS) |
|---|---|
| State | forecasted prices (DAM → BM), physical position & limits, unit availability, forecast error distributions |
| Decision | submit price-quantity curves in DAM/IDA, adjust PNs, dispatch schedules, charge/discharge set-points, imbalance hedges |
| Exogenous info | actual cleared prices, TSO dispatch instructions, imbalance prices, real-time production, network constraints |
| Transition | battery SoC dynamics, forecast updates, merit-order changes, rule-driven position settlement |
| Objective | maximise expected gross margin (energy + system-service revenue) subject to risk or reserve constraints |
The trader therefore faces a sequential stochastic optimisation problem with partially observable dynamics and multiple market layers.
3 I-SEM market structure and timing constraints
The Irish SEM currently offers the following short-term trading levers:
- Day-Ahead Market (DAM) – closes 11:00 D-1; establishes baseline position.
- Intraday Auctions – SEMOpx operates three auctions:
- IDA-1 17:30 D-1, covers full delivery day
- IDA-2 08:00 D, covers 11:00–23:00
- IDA-3 14:00 D, covers 17:00–23:00 (SEMOpx)
- Intraday Continuous – order book remains open until 60 min before real-time (inter-connector trades until 30 min).
- Balancing Market (BM) – dispatchable units must submit Physical Notifications and CODs by gate closure 13:30 D-1 and may update them up to the Gate Closure of each half-hour imbalance period (cms.soni.ltd.uk).
These windows generate a nested decision sequence: DAM → IDA-1 → IDA-2 → IDA-3 → Continuous → BM, each step revising earlier commitments under uncertainty about renewable output, demand and imbalance prices.
4 Analogous domains that inspire solution methods
- Chess / Go – large combinatorial state spaces tackled with tree search + learned value networks.
- Robotic path-planning & self-driving – continuous control under stochastic disturbances, solved with model-predictive control and reinforcement learning.
- Real-time ad-bidding – sequential budget allocation with censored demand observations, often addressed via bandit algorithms.
Each field leverages look-ahead, value function approximation or policy search – the same meta-classes Powell defines for energy trading.
5 Implications for I-SEM optimisation tools
- Comprehensive state representation – include forecast distributions, system-service opportunity costs and unit constraints.
- Rolling horizon decision analytics – recalculate actions at every market gate using updated beliefs.
- Learning layers – continuously update error distributions and price impacts; treat forecast models as evolving state components.
- Risk-aware objective functions – embed downside metrics (e.g., Conditional VaR of imbalance charges) directly in the contribution function.
Such structure enables transparent conversation with regulators and investors while paving the way for algorithmic policies (PFAs, CFAs, VFAs, DLAs) chosen according to computational tractability and data availability.
References
- Powell, W.B., Sequential Decision Analytics and Modeling: Modeling with Python, Foundations & Trends®, 2022
- SEMOpx, "Intraday Auctions Market – How it works", accessed 30 Apr 2025 (SEMOpx)
- EirGrid & SONI, Balancing Market Principles Statement V8.0, pp. 33–36, 2024 (cms.soni.ltd.uk)
In the next article, we will examine methodological options – from rule-based heuristics to reinforcement-learning agents – that address this decision structure while respecting I-SEM's operational and regulatory constraints.