When CSI 300 rallies, most stocks still lose

This note is educational research, not investment advice. It is a research prototype, not a live track record and not a recommendation to trade.

Most people who want to try systematic work start with one question: can an idea make money? The honest version is narrower — on paper, after costs, versus just holding the index, and how badly it hurts along the way.

CSI 300 is cap-weighted. A handful of large growth names can lift the index while the typical stock goes nowhere. We built a defensive A-share book around that puzzle: prefer quiet names, lean cheap, and cut exposure when the rally is narrow and the book is already bleeding.

Window: 10 October 2016 to 21 August 2026, after commission, stamp tax, and 0.1% slippage. The next cell is the snapshot that answers that question.

Did it make money?

On this paper tape: yes, and it was quieter — with one stretch where it looked late on purpose.

• Over ten years the prototype compounded to about +117%. CSI 300 did +40%. The worst peak-to-trough drop was almost halved (−26% vs −46%).
• It is not a straight line. The book lost less in 2018. It made money in 2021–23 while the index fell. In 2019–20 a narrow growth rally left cheap, quiet names behind. That lag is the design, not a surprise bug.
• After that failure showed up on the tape, we added two brakes: a six-month style cut, and a breadth-plus-drawdown cut. Same idea as a weather check before you size the book.

If you already run research for a living, the rest of this note is a reusable loop — write the idea, cost it, log the year it failed, add a brake — so the next prototype is not another one-off script.

Start here: a safe, small loop

You do not need a trading desk, a PhD, or a real-time feed to learn this kind of work. You need a loop you can repeat without hurting yourself.

Prepare four things:

1. One sentence. Write the idea before you code. Ours: overweight quiet, cheaper A-shares; cut when the index rally is narrow and the book is already down.
2. Dated data. Daily prices for a stable universe, plus earnings that use announcement dates. Wrong dates make a beautiful, false backtest.
3. Costs in the model. Commission, stamp tax, slippage — in the same units as returns. Rebalance slowly (we use January and July) so turnover does not eat the idea.
4. A slice you do not tune on. If the later years look better than the early years, treat that as a warning light, not a trophy.

Stay on paper until that loop is boring. Live money is a later tool, not an entry ticket.

Tools that are enough. Python and a notebook will run a first prototype. What experienced researchers usually need is not a fancier oscillator — it is one project that keeps data, code, charts, and a short written log together, so every new idea is not copy-pasted across chat windows and a fresh blank file.

The rest of this note applies that loop to this A-share book.

Keep going: the recipe, the brakes, the year it failed

You now have the loop. Subscribe free to see it applied to this book:

• the three jobs in plain language, then short code you can read without a quant background
• how the brakes fire only when the failure mode is already on the tape
• 2019–20, when CSI 300 told the truth and the book did not listen
• what we still do not trust — so you do not copy a trophy Sharpe

Subscribe free to read it — no fees, research only.