JSE Portfolio Risk: Out-of-Sample Tests

Optimised portfolios and risk numbers look precise when they are estimated and judged on the same data. The original app chose weights with hindsight, compounded log returns as if they were simple returns, and measured shortfall against a target rather than a loss, so its outputs said little about how a portfolio would actually behave.

Every rule now rebalances at each month end using only prices up to that day, holds until the next month end while weights drift, and pays 10 bps a side plus 0.25% securities transfer tax on purchases. The universe is the original app's 267 tickers, narrowed each month to the 40 most traded. The design period chose each rule's lookback; the validation period was run once, and Sharpe ratio differences were tested with a HAC test rather than read off the table. A Python service on Azure refreshes prices from Yahoo Finance, extends the live record and fits portfolios on request.

• Rules: the original random-weights method, equal weight, inverse volatility, minimum variance, equal risk contribution and maximum Sharpe, with Ledoit-Wolf covariance and a 10% cap on any share
• Benchmarked against the Satrix 40 ETF, with the three-month Treasury bill rate from FRED as the risk-free rate
• One-day 99% and 97.5% VaR and expected shortfall from historical simulation, a normal model, EWMA and filtered historical simulation on a GJR-GARCH model with Student t errors refitted every month
• VaR backtests: Kupiec unconditional coverage, Christoffersen independence and conditional coverage, the Basel traffic light on rolling 250-day windows, and the Acerbi-Szekely Z2 test for expected shortfall
• Stress tests: the 2008 crisis, December 2015, Covid and the April 2025 tariff shock replayed on current holdings, plus correlations at 0.9, doubled volatility, a 20% fall in the rand and the largest holding going to zero
• A monthly job downloads fresh prices, reruns the pipeline and redeploys the dashboard

• Python research pipeline writes the backtests, risk-model tests and stress results as JSON
• FastAPI service shares the same library, serves the live record since 1 October 2026 and runs the portfolio builder on Azure App Service
• React dashboard in this site's typography and palette, embedded here and usable on its own

• Minimum variance did what it is meant to: 14.8% annual volatility against 18.5% for equal weight and 18.8% for the Satrix 40. It also returned −0.6% a year, and its Sharpe ratio was significantly below equal weight's (t = −2.94)
• Maximum Sharpe was the best rule, 8.4% a year with a Sharpe ratio of 0.18 against 0.09 for equal weight, but the difference is indistinguishable from noise (t = 0.42)
• The original app's method returned 3.1% a year (Sharpe −0.09), significantly worse than equal weight (t = −2.36), with about six times the turnover
• No rule beat the Satrix 40, which returned 9.3% a year (Sharpe 0.22)
• Only filtered historical simulation passed the VaR and ES backtests, for all three portfolios tested: 25 breaches of the 99% VaR for equal weight against 24 expected. The normal, EWMA and plain historical models all failed, mostly by breaching too often and in clusters
• Live figures since 1 October 2026 are in the preview. With weeks of unseen data they are not yet a meaningful test

• All results are backtests or a live paper record, not realised performance or investment advice.
• The universe is today's list of tickers, so shares that delisted earlier are missing. This survivorship bias flatters every rule.