{"paper_id":"2308.06935","title":"Insurance pricing on price comparison websites via reinforcement learning","authors":["Tanut Treetanthiploet","Yufei Zhang","Lukasz Szpruch","Isaac Bowers-Barnard","Henrietta Ridley","James Hickey","Chris Pearce"],"paper_date":"2023-08-14","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2308.06935v1","page_url":"https://thequant.space/flowcharts/insurance-pricing-on-price-comparison-websites-via-reinforce/","scores":{"empirical_rigor":6.5,"math_complexity":6,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes a hybrid reinforcement learning framework that combines model‑based offline training with model‑free contextual bandit updates to learn optimal insurance pricing policies on price comparison websites.","quote":"To address these challenges, this paper introduces reinforcement learning (RL) framework that learns the optimal pricing policy by integrating model-based and model-free methods.","summary":"The emergence of price comparison websites (PCWs) has presented insurers with unique challenges in formulating effective pricing strategies. Operating on PCWs requires insurers to strike a delicate balance between competitive premiums and profitability, amidst obstacles such as low historical…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"Customer features (X_t) are independently and identically distributed.","type":"statistical","stated":"explicit","quote":"To facilitate training an agent, we make the following assumptions on the random variables (Xt, Yt)T t=1 in (2.1): (i) The customer features (Xt)T t=1 are independently and identically distributed.","support":"full"},{"assumption":"The market quantile price(s) H_t follow a conditional distribution ψ_h(·|X_t).","type":"market","stated":"explicit","quote":"(ii) For each t ∈{1, . . . , T}, the market quantile price(s) Ht of all agents takes values in a space H, and follows the (conditional) distribution ψh(·|Xt).","support":"full"},{"assumption":"The customer’s decision Y_t is a conditional Bernoulli random variable.","type":"model","stated":"explicit","quote":"(iii) For each t ∈{1, . . . , T}, the customer’s decision Yt is a conditional Bernoulli random variable.","support":"full"},{"assumption":"The conversion probability p(X_t, H_t, A_t) is a deterministic function independent of time t.","type":"model","stated":"explicit","quote":"Then Yt = 1 with probability p(Xt, Ht, At) and 0 otherwise, where p : X × A × H →[0, 1] is a deterministic function independent of t.","support":"full"},{"assumption":"A time‑independent pricing policy ϕ*: X → A maximises the one‑step reward, allowing the problem to be reduced to a contextual bandit.","type":"model","stated":"explicit","quote":"These modelling assumptions imply that it suffices to find a time-independent policy ϕ⋆: X → A that maximises the one-step reward: ϕ⋆∈arg min ϕ:X→A E [Yt · r(Xt, ϕ(Xt))] .","support":"full"}]},"datasets":{"status":"available","items":[{"name":"Synthetic dataset of 35,000 customers","provider":null,"kind":"synthetic","access":"proprietary","period":null,"universe":null,"quote":"4 Numerical experiments To demonstrate our methdology, we generate synthetic dataset that is reflective of a private commercial PCW data made available to the authors1. A synthetic dataset of 35000 customers on a PCW, with 16 customer features (spanning age to credit risk data) along with simulated Avg.","support":"full"}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"synthetic"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R4a","candidates":{"front_matter":"n/a","evidence_card":"yes","evidence_card_quoted":true},"card_detail":"Burn cost b(x) is modelled as b(x) = Avg.Top5 × N(0.8, 0.2) and is subtracted from the reward function.","card_quote":"Comparing to real data, we found the following calculation reflected these real-world con-straints: b(x) = Avg. Top5 × N(0.8, 0.2), P0(x) = Avg.","meaning":"the paper reports no trading or execution result; the cost the extractor found is a model parameter"},"results":{"status":"available","evaluation_type":"none","findings":[{"finding":"The hybrid RL agent outperforms all benchmark agents—including the standard model‑free RL agent—except an unrealistic perfect‑information agent.","kind":"performance","quote":"Although the proposed hybrid RL (brown line) underperforms the (unrealistic) perfect information agent, it clearly outperforms the remaining 5 agents, including the standard model-free RL agent (purple line).","support":"full"},{"finding":"The hybrid agent accumulates reward faster than all other agents, demonstrating improved sample efficiency.","kind":"performance","quote":"Furthermore, we observe the rate at which the hybrid agent accumulates reward outperforms all other agents, again with the exception of the perfect information agent.","support":"full"},{"finding":"The hybrid approach shows higher cumulative expected and realised rewards compared to traditional pricing strategies.","kind":"performance","quote":"This demonstrates the improved sample efficiency from use of our hybrid approach compared to more traditional pricing strategies/benchmarks.","support":"full"}],"reported_performance":null,"out_of_sample":{"method":"holdout","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":9,"citations":{"cited_by_in_corpus":0,"by_independent_authors":0,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":false,"code_url":null,"note":"No repository presented by the authors as their own was found in the paper. Absence here is not proof that no code exists."},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":2,"extracted_at":"2026-11-14T20:01:39Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}