Portfolio Optimisation

A brute-force portfolio grid search over long-run daily asset histories. It tests every weight combination at a fixed increment, backtests each one from 1970 to the present, and writes a ranked results table plus a self-contained interactive chart you can open in any browser.

The point is to see the whole trade-off surface at once rather than a single "optimal" answer. Every blend is a dot, so you can look at what growth actually costs in drawdown, where the diminishing returns start, and how wide the band of near-equivalent portfolios really is.

Above: every 5% blend of four assets over 1970 to 2026. The pinned dot is the shallow-drawdown corner, 20/60/15/5 across stocks, bonds, gold and commodities, which gave up about two points of CAGR against the best blends but fell only 21% at its worst instead of 40%+. The curved upper edge is the frontier, and the width of the cloud beneath it shows how many quite different recipes land in much the same place.

pip install pandas numpy
python portfolio_optimizer.py

Outputs land in output/:

File	What it is
portfolio_optimisation_results.csv	Every portfolio tested, ranked, with all metrics
portfolio_optimisation_results.html	Interactive chart and table, no dependencies, no server needed

A generated example of both is committed to this repo. The screenshot above is a still of that page; download the HTML and open it locally to get the interactive version, since GitHub will not render it in the browser.

Data

Price data comes from the companion repository mwilczynska/financial_datasets. Every series there runs from 1970 to the present. No real fund goes back that far, so only the most recent stretch of each series is a fund's own returns; before that it is reconstructed from published index data and documented models, with the method and its limits written up per asset. Check the repo out as a sibling directory:

your-folder/
├── financial_datasets/
└── portfolio_optimisation/

DATA_ROOT in portfolio_optimizer.py resolves to ../financial_datasets/data/processed. Point it somewhere else if your layout differs.

Two asset sets are wired up out of the box. The default is un-levered:

Code	What it is	Series covers
GLSTOCK	All-world equities in USD, dividends reinvested (VT-like)	1970 on, VT's own returns from 2008
GLBOND	Unhedged global government bonds in USD (BND/BWX blend)	1970 on, the fund blend from 2007
GOLDPM	Gold bullion as a total return (GLD-like)	1970 on, GLD's own returns from 2004
COMM	Broad commodity futures with roll yield and collateral (DBC-like)	1970 on, DBC's own returns from 2006

The alternative, commented out in the config block, is a leveraged set: USLCAP3x, LTT3x, ITT3x, GOLDPM2x and COMM, modelling daily-reset funds like UPRO, TMF, TYD and UGL with their financing costs, fees and volatility decay included. Those series also start in 1970 and switch to the real funds when they launched, around 2008 to 2009. The generated page adapts its explanations and warnings to whichever set you actually ran.

Configuration

Everything editable sits at the top of portfolio_optimizer.py:

Setting	Purpose
asset_list	Which assets to blend
weight_step	Grid increment. 0.05 for the full 5% run, 0.20 for a quick test
min_weights / max_weights	Per-asset floors and ceilings
max_drawdown_limit	Drop portfolios worse than this, or None to keep everything
rank_by / rank_ascending	Which metric sorts the table
rebalance_annually	Annual rebalance, or buy and hold
starting_value	Opening pot for every portfolio
risk_free_rate_annual	Used by Sharpe and Sortino

Grid size grows fast: four assets at 5% steps is 1,771 portfolios, five assets is 10,626.

Metrics

Final value, CAGR, annual return, standard deviation, TWRR, Sharpe, Sortino, Calmar, max drawdown, Ulcer Index, average drawdown, longest drawdown in days, daily win rate, best and worst day, month and year, 95% VaR and CVaR, skew, and excess kurtosis.

Each row also carries an Existing Script Syntax column, a ready-to-paste tickers, weights = [...],[...] line for use in other backtesting scripts.

The interactive page

Written as plain HTML, SVG and JavaScript with no libraries and no build step, so it works from a local file or any static host.

• Pick what the X and Y axes measure from any numeric column.
• Hover a dot for that portfolio's mix and statistics. Click to pin it, click again to unpin.
• Highlight every dot past a threshold you choose on the current X axis.
• Sort the table by any column, search the weights, or stack numeric filters such as "Max Drawdown at most 60%". Loss-type columns are magnitude-aware, so that filter keeps the shallower drawdowns rather than matching everything.
• Chart and table are linked. Pinning a dot scrolls to and highlights its row, and vice versa.

The page also carries a plain-language explainer of what it shows, how each data series was built, and what to be sceptical about. ABOUT.md is a standalone copy of that write-up.

Caveats

This is a historical simulation, not advice, and past results do not predict future returns. Hindsight is baked into any ranking of backtests: the top row is the blend most fitted to this particular history. Taxes, commissions, bid/ask spreads and rebalancing friction are not modelled. The early decades of every series are reconstructions rather than observed history, since no real fund reaches back to 1970, and ABOUT.md spells out which parts, for which asset.

Licence

Code is MIT, see LICENSE.

The datasets are not covered by that licence. Some upstream sources carry their own redistribution or non-commercial terms, notably the S&P GSCI republication behind the early commodity history and the Jorda-Schularick-Taylor macrohistory database behind the early bond history. See DATA_LICENSE.md in mwilczynska/financial_datasets before reusing the data itself.