Reinforcement Learning
Reinforcement learning (RL) is training a system by rewarding the outcomes of its actions rather than by showing it correct answers: an agent acts in an environment, receives a reward, and adjusts its policy — its mapping from situations to actions — to earn more reward over time. It produced the game-playing systems (AlphaGo, 2016), it is the RL in RLHF that shaped chat models, and it trained the reasoning models of 2024–2025; in fintech it optimises pricing and collections.
- Also known as
- RL, Reinforcement learning (RL), Reward-based learning, Policy learning
- Used at
- games and simulations, robotics, recommendation and advertising (bandits), dynamic pricing, marketing and collections strategy, order execution and trading, resource scheduling, and, in AI development, the post-training of language models (RLHF) and reasoning models (RL with verifiable rewards on maths and code)
- Binding
- none for the method; systems that act on customers fall under the sector rules on those actions (pricing fairness, collections conduct, marketing consent) and the EU AI Act by use; the alignment concern — a system optimising a reward the designers specified imperfectly — is the field's central risk and the motivation for much AI-safety work
- Typical mechanics
- a policy (often a neural network) chooses actions; the environment returns a state and a reward; algorithms (Q-learning, policy gradients, PPO, and the bandit algorithms for one-step problems) update the policy toward higher expected reward, balancing exploration (trying actions to learn) against exploitation (using what is known); training needs many episodes, which is why simulators and offline data matter
- Typical difficulties
- reward specification (the system optimises what is rewarded, not what was meant), sample inefficiency (millions of trials), exploration cost when trials are real (a mis-priced product, a wrong collections action), instability, and evaluation (a policy's value is what it does over time, not a single accuracy)
- Parties
- the researchers (DeepMind, OpenAI and the academic groups), the labs using it for post-training, the practitioners in marketing and pricing, the trading firms, and the safety researchers whose field it partly defines
How reinforcement learning works in practice
A lender wants to choose, for each delinquent customer, which contact strategy to use — SMS, call, email, a payment-plan offer, or wait — to maximise recovery within conduct rules. Supervised learning would need labels for the right action, which nobody has. A contextual bandit, the simplest RL, treats each customer's features as the context, each strategy as an action, and the recovered amount minus cost minus a penalty for complaints as the reward; it tries strategies with controlled exploration, learns which work for whom, and shifts toward them. After three months it recovers 8% more than the fixed policy at the same complaint rate. The exploration was the cost: 10% of customers received a strategy chosen partly at random, within the rules, so the system could learn.
The larger uses are in the AI labs. RLHF trains a language model to produce responses a reward model (trained on human preferences) scores highly; the reasoning models of 2024–2025 were trained with RL on problems with verifiable answers — maths, code — rewarded for getting them right, and learned to produce long chains of reasoning on their own. Both show the method's power and its risk: a policy optimises the reward it is given, and every gap between the reward and the intent — sycophancy that raters preferred, a chain that reaches the answer by a route nobody wanted — is exploited. For a fintech, RL is a tool for sequential decisions where outcomes can be measured and exploration is affordable, with the reward specified by people who understand what the system will do to maximise it; trading is the domain where the reward is money and the environment adapts, which is why the successes there are narrow and guarded.
For a fintech, reinforcement learning is the method for decisions that are strategies, not predictions — with the reward written by the compliance team as much as the data team; for an investor, it is the technology behind the reasoning models and the alignment problem in one.
Example
A payments company uses reinforcement learning to choose retry timing and routing for failed subscription payments: for each failed charge, the actions are when to retry (hours to days) and through which acquirer; the reward is recovered revenue minus fees minus a penalty for each retry (issuers penalise excessive attempts). The policy is trained offline on two years of retry logs, then deployed with 5% exploration. Results after six months: recovery rate up from 62% to 71% of failed charges, retries per recovered charge down from 3.1 to 2.2, scheme penalties down 40% — €5 million of recovered revenue a year. The constraint that mattered: the reward's penalty term was set by the payments team from the scheme rules, because without it the policy learned to retry constantly. The policy's actions, rewards and exploration rate are logged and reviewed monthly.
Common mistakes
- A reward that omits a cost, and a policy that exploits the omission. - Exploration on real customers without limits set by conduct rules. - Treating an RL system's single-run accuracy as its value; the policy's long-run behaviour is. - RL where supervised learning on existing labels would do. - Trading systems that learned a regime and meet the next one.