RL Alpha Labs

Judged after real costs and tax, not by the best backtest. RL Alpha Labs does automated trading research on Indian stocks. This repository is its website, not code: the trading code is private. Each idea is tested on data from NSE, India's National Stock Exchange, with the companies that later failed kept in, and must beat an index fund or a simpler method after real costs and tax. No live capital is deployed.

Did any idea work?

Most trading ideas fail, the lab's notes say; the useful part is knowing why. One that closed: holding 300 stocks yourself instead of an index fund. In a 20.7-year backtest (2006 to 2026), rebalancing them every quarter lost 2.14% a year to trading costs and yearly capital-gains tax, and trailed the index fund in 69 of 70 combinations of account size and number of stocks. The cause is built in: a fund rebalances inside itself, so its investor pays the tax once, at the end.

Twelve ideas are closed, each with its reason: costs and taxes closed four; five showed no edge or only luck; in one, prices rose before the news, not after; two AI models missed the bar set for them.

Start with the eight backtest traps that fooled us →

What one trade costs

Many backtests assume 10 to 15 basis points a trade. A buy and a later sale on NSE cost this, from the lab's costs note (rates read on 22 October 2026; a basis point is 0.01%):

Cost	Basis points
Taxes and exchange fees	22.25
A flat ₹15.34 fee for each stock sold, on a ₹5,000 position	30.68
The same fee on a ₹50,000 position	3.07
The spread between buying and selling prices, large to smaller companies	24 to 55.4

The eight traps

Each made a result look better, or a check look passed, without an error message:

1. Trading at the price you decided on
2. Testing only the companies that survived (roughly 2.4 to 4.2 points a year, in the lab's estimate)
3. Prices of zero at the start of history
4. Dropping what you cannot match
5. A pass mark mistaken for power
6. A control that cannot fail
7. A broken measurement that looks like progress
8. A language model that already knows the answer

What it is not

• Not a product or advice. No signals and no managed money. RL Alpha Labs is not registered with SEBI as an investment adviser or research analyst; nothing here is investment advice.
• Not live. Simulation and paper trading (orders recorded, never sent) only. No live capital is deployed, and no AI model places trades.

How it works

The stages are in Methodology.

Where next

• Open problems: twenty research problems the lab has not solved, open to collaborators.
• Join the lab: what working with it looks like, and the open projects.
• What is in this repository: the site's files, and how to preview a change on your own machine.
• Other tools from the same lab: Servedeck, GLM-5.3 on one GPU and hostops.

---

_{Big models, small places, real numbers · RL Alpha Labs is built by pctablet505. Past simulation results do not guarantee future live performance.}