Investment management technology is several distinct disciplines that all must perfectly converge to deliver value to the end user:
Portfolio analytics. Position-level and aggregate exposure, factor loadings, beta and alpha decomposition, sensitivities at the book level, concentration and style/sector/geography slicing, liquidity profile, pre-trade and post-trade what-if analysis. The layer the PM actually looks at.
Portfolio construction and rebalancing analytics. Mandate-aware optimization (mean-variance, risk-parity, hierarchical risk parity, CVaR, Black-Litterman with custom views), transaction-cost-aware rebalancing, drift monitoring against model portfolios, tax-aware lot selection, trade-list generation with minimum-ticket and round-lot constraints, integrated with pre-trade compliance and broker-aware cost models.
The front office. OMS and EMS, FIX connectivity to brokers, venues, and dark pools, smart order routing, algo wheels, and the pre-trade compliance layer that has to say “no” in milliseconds. Hard problems: latency, correct state under partial fills, modelling the trader’s intent.
The middle office. IBOR keeps the portfolio as the PM sees it; ABOR keeps it as the auditors and fund admin do. They disagree on purpose — trade date vs. settle date, economic vs. legal ownership, local vs. base currency, tax-lot treatment — and reconciling them cleanly is where most operational risk hides. Real-time P&L, greeks, and exposure live here, alongside post-trade compliance, restricted lists, 1940 Act, UCITS, Reg-T, portfolio margin, leverage, liquidity, and counterparty limits.
The back office. Performance first: GIPS-compliant composite construction and maintenance, time-weighted and money-weighted returns, and Brinson / Brinson-Fachler attribution extended for fixed-income effects (curve, spread, carry, roll-down, selection). Then the plumbing underneath: reconciliation against prime brokers, custodians, and fund administrators; corporate actions; security master; pricing waterfalls; collateral; client reporting. Unforgiving: every mismatched position or stale price eventually becomes a restated NAV.
The asset-class surface. Equities (common, preferred, ADR, ETF); listed and OTC options (vanilla, exotic, structured) on a working vol surface; futures and futures options with margin mechanics; FX spot, forward, NDF, and option; cash fixed income (UST, TIPS, agencies, IG/HY corporates, munis, sovereigns, EMD) with curve and spread conventions; securitized credit (agency and non-agency MBS, CMBS, ABS, CLO, syndicated loans) with prepayment and default modelling; credit derivatives (single-name and index CDS, CDX, iTraxx, tranches) with hazard-rate and copula models; interest-rate derivatives (IRS, OIS, basis, swaptions, caps, floors, FRAs) on a multi-curve framework and vol cube; inflation (ZC and YoY); total-return swaps; repo; crypto spot and perps. The interesting bugs live where two of these meet.
Single-security analytics. The greek and sensitivity stack — delta, gamma, vega, theta, rho, DV01, key-rate duration, convexity, OAS, spread duration, prepay-adjusted yield — plus term-structure models (Hull-White, LMM, SABR), credit models (hazard-rate, structural, copula), and Monte Carlo / PDE / closed-form pricers where each is appropriate. The art is knowing which approximations a pre-trade screen tolerates and a risk report does not.
Portfolio-level risk. Historical, parametric, and Monte Carlo VaR; expected shortfall; marginal and component contribution; incremental and reverse stress; factor decomposition (Barra-style and custom a-priori); tracking error, information ratio, Sharpe, Sortino, Calmar, max drawdown, beta, alpha. Each answers a different question; conflating them is how risk reports quietly mislead.
Performance attribution. Equity attribution is mostly solved by Brinson-Fachler with sector or style buckets. Fixed income is the real work: decomposing return into curve, spread, carry, roll-down, and selection within sector/rating/duration buckets while keeping the arithmetic additive and the residual near zero. Multi-currency and multi-period linking add further wrinkles.
Problems worth building for.
- Scenario and stress testing with parallel and non-parallel shifts across benchmark, curve, vol surface, spread, FX, and correlation regimes — historical replays (’87, ’98, ’08, ’20) and custom.
- Real-time liquidity risk: time-to-liquidate curves, market-impact cost (Almgren-Chriss), intraday funding and haircut alerts.
- Fund- and firm-level capitalization and margin stress: Reg-T, TIMS, STANS, OCC portfolio margin, SPAN, and house methodologies. (Three implementations FINRA-approved at separate broker-dealers.)
- Open, tunable a-priori factor models where vendor factors don’t fit the mandate.
- Portfolio construction and optimization: mean-variance, risk-parity, hierarchical risk parity, CVaR, Black-Litterman, transaction-cost-aware.
- Data infrastructure: security master, corporate actions, pricing waterfalls, golden-copy delivery from Bloomberg, Refinitiv, ICE, S&P/IHS, MarkIt, FactSet, and fund admins.
Where AI changes the picture. ML and LLMs are reshaping benchmark approximation and sparse portfolio construction, cross-sectional and time-series momentum, optimal block execution and VWAP/TWAP scheduling, margin and loss prediction, sentiment from filings and news, research triage, and automated reconciliation. The question is rarely “can a model learn this?” — it is “what data, loss function, and guardrails make the answer trustable inside an investment process?”
Experience. Six production investment-management platforms shipped across three decades — OMS/PMS, analytics, risk, and performance — at firms from boutique buy-side shops to multi-vendor fintech. Two were sold; one still runs on the original code. Roles from Director of Research and founder through C-level and executive management, down to writing and debugging code on the desk this afternoon. Still hands-on.
Teams. Investment-management technology rewards small teams of exceptional people over large teams of average ones. The systems are too tightly coupled for coordination overhead to pay for itself.
Technology stack. Python (NumPy, Pandas, Polars, PyTorch, scikit-learn, statsmodels, QuantLib), Java and Node.js for low-latency services, C# / .NET on the buy-side desktop, SQL (MSSQL, PostgreSQL, kdb+/q), FIX 4.x / 5.x, REST, gRPC, Kafka, Redis, Linux.
