Five sections covering the structural decisions, the data model under them, and the operational characteristics you can show to an architecture review or a regulator.
§ 01
Why mean-variance is not enough on its own
Markowitz mean-variance optimisation is the foundational allocation framework, but applied naively it produces extreme corner solutions: 100% concentration in whichever asset has the highest historical Sharpe, near-zero allocations to anything else. The reason is that mean-variance is hyper-sensitive to estimated returns, and estimated returns over short windows have low signal-to-noise. Coreal's allocator uses mean-variance for portfolio shape but applies regularisation (shrinkage, robust covariance, allocation bounds) to suppress noise-driven concentration. The result is a portfolio that captures the mean-variance benefit without the brittleness.
§ 02
Black-Litterman: blending market views with priors
Black-Litterman lets the system express informed views ("emerging-market debt is undervalued by 100bps relative to consensus") as priors blended into the optimisation. The mathematical framework is well-established; the engineering challenge is sourcing views responsibly. Coreal's view-injection layer takes views only from explicit, journaled sources (research provider feeds, declared house views, customer-declared values like "ESG-aligned"), with the source and its confidence surfaced in the recommendation. A user sees the resulting portfolio with an explanation: "your target reflects views from sources X, Y, Z, weighted by Q". No black-box opinion injection.
§ 03
Suitability gating at two levels
Suitability is enforced at both the portfolio level (does this overall allocation match the user's risk tolerance and goals?) and the position level (is this individual instrument appropriate for this user, regardless of weight?). Position-level gating catches cases where an aggregate-suitable portfolio contains individually-unsuitable instruments — leveraged ETFs, derivative structures, single-stock concentrations beyond limits, instruments restricted in the user's jurisdiction. Both levels are checked at portfolio construction and at every rebalance. Failures surface to the user with a plain explanation; a regulator audit can replay the exact suitability decision for any instrument the user ever held.
§ 04
Goal-based planning with Monte Carlo
Many roboadvisers offer "goal-based" planning that reduces to a glide-path: "if you're saving for retirement in 30 years, here's an aggressive allocation now, gradually shifting to bonds". This is a model, not a plan. Coreal's goal planner runs Monte Carlo simulations over the actual portfolio against user-defined goals (timing, amount, priority), reporting the success probability per goal and the marginal impact of contribution changes. When a user asks "what happens if I add £200/month?", the answer is a quantified delta in success probability for each goal, not a slogan.
§ 05
Drift, rebalance, and tax-aware execution
Portfolios drift from target as markets move. Rebalancing brings them back, but mechanical rebalancing produces unnecessary tax events (in taxable accounts) and trading costs. Coreal's rebalance engine combines threshold triggers (rebalance only when drift exceeds N percentage points) with tax-aware execution (sell highest-cost-basis lots first, prefer rebalancing through new contributions, defer realising losses where wash-sale rules apply). Every rebalance proposal is shown to the user with an estimated tax impact before execution.