{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,9,27]],"date-time":"2026-09-27T13:57:49Z","timestamp":1790517469070,"version":"4.1.0"},"posted":{"date-parts":[[2026]]},"group-title":"SSRN","reference-count":0,"publisher":"Elsevier BV","license":[{"start":{"date-parts":[[2026,1,1]],"date-time":"2026-01-01T00:00:00Z","timestamp":1767225600000},"content-version":"unspecified","delay-in-days":0,"URL":"https:\/\/www.uspto.gov\/ip-policy\/copyright-policy\/copyright-basics"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"abstract":"<jats:p>Reinforcement-learning (RL) portfolio agents trained for developed marketsgeneralize poorly to emerging and frontier markets, where returns arenon-stationary, transaction costs high, and policy rates can move several foldwithin a single year. We study two questions across a six-market panel, namelySri Lanka, Egypt, Argentina, Brazil, China, and Indonesia, spanning frontier andemerging tiers and macro-fundamental, hyperinflationary, and flow-driven crises.First (RQ1), replacing the reward&amp;apos;s fixed risk-free benchmark with the country&amp;apos;scontemporaneous policy rate, together with a conditional drawdown penalty,produces a reward whose defining and universal signature is maximum-drawdowncontrol: across all six markets and both bearish and bullish regimes, that istwelve of twelve test windows evaluated over five seeds, the dynamic rewardyields materially shallower drawdowns than a static Sharpe reward. In the crisesthe reduction is large, for example 14.6\\% versus 35.6\\% in Egypt, 12.9\\% versus30.8\\% in China, and 27.3\\% versus 38.9\\% in Sri Lanka, and the agent&amp;apos;s drawdownis shallower than passive buy-and-hold in every market. Because the propertyholds even where the local rate is flat, we isolate the drawdown penalty ratherthan the moving benchmark as the mechanism. Second (RQ2), we introduce a CrisisStructural Similarity (CSS) metric, a kernel two-sample distance between amarket&amp;apos;s crisis-window macro-stress features and those of US crises, and showthat it anticipates whether transfer reduces training-outcome variance (Spearmanrank correlation of 0.77 with the crisis-window stability gain). A similaritythreshold separates the three stabilizing markets from the non-stabilizing onesand is confirmed out of sample on a held-out market, whose predicted-neutralstability gain was measured at 1.00. Transfer therefore behaves as a conditionalregularizer whose benefit is predictable a priori from crisis structure ratherthan discovered post hoc. The work delivers a reward design with a robust,quantifiable risk-control property across heterogeneous markets, and atransferability criterion that indicates in advance whether cross-marketpre-training is worthwhile.<\/jats:p>","DOI":"10.2139\/ssrn.7532406","type":"posted-content","created":{"date-parts":[[2026,9,27]],"date-time":"2026-09-27T13:37:20Z","timestamp":1790516240000},"source":"Crossref","is-referenced-by-count":0,"title":["Dynamic Risk-Free Benchmarking for Robust Drawdown Control in Reinforcement-Learning Portfolio Management: Evidence from Six Emerging and Frontier Markets, with a Crisis-Similarity Criterion for Cross-Market Transfer"],"prefix":"10.2139","author":[{"given":"Kalindu  Binal","family":"Weerasena","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Y.  H. P. P.","family":"Priyadarshana","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"78","container-title":[],"original-title":[],"deposited":{"date-parts":[[2026,9,27]],"date-time":"2026-09-27T13:37:20Z","timestamp":1790516240000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.ssrn.com\/abstract=7532406"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026]]},"references-count":0,"URL":"https:\/\/doi.org\/10.2139\/ssrn.7532406","relation":{},"subject":[],"published":{"date-parts":[[2026]]},"subtype":"preprint"}}