An Open-Source Backtesting Stack for $0

Backtests lie politely when the stack is sloppy. Free tools can still be rigorous if you treat them like field gear: labeled, redundant, and never trusted blindly. Here is a surplus parts manifest for a zero-dollar research stack that is good enough to reject bad ideas quickly — which is most of the job.

Parts manifest

1. Data — Free EOD equity feeds (Polygon free, FMP free, or carefully licensed community sources). EDGAR for filings. Accept delay; do not pretend free tiers are live.
2. Wrangling — pandas / Polars in a local notebook. Cache downloads. Log the as-of date on every CSV.
3. Indicators — pandas-ta, TA-Lib, or a small hand-rolled set you understand. Prefer fewer indicators you can explain.
4. Engine — A vectorized or event-driven open backtester with explicit commissions, slippage, and position sizing hooks.
5. Diagnostics — Drawdown, turnover, hit rate, and a simple walk-forward split. If the library cannot report costs, add them yourself.
6. Rehearsal — Ideas that survive the notebook still need human execution practice on Stock Picks.

Assembly order

Start with one liquid universe (for example large-cap US) and one boring strategy (SMA cross or RSI threshold). Wire data → signals → positions → P/L with a fixed dollar risk per trade. Only after that pipeline is boring should you add complexity.

Common failure modes on free stacks:

• Survivorship-biased universes (missing delistings)
• Ignoring borrow and locate for short legs
• Optimizing twenty parameters on five years of data
• Using adjusted closes incorrectly for corporate actions
• Forgetting that free API rate limits truncate history mid-test

Document each known limitation in a LIMITATIONS.md next to the notebook. Future you will thank present you.

Walk-forward or it did not happen

A single in-sample equity curve is a story, not evidence. Split time: train on an earlier window, validate on a later window, and keep a final holdout you rarely touch. If the edge vanishes out of sample, the stack worked — it saved you from funding a myth.

Grid search is allowed only with a tight parameter budget and a cost model. Unlimited search on free data is how curve-fitters go broke slowly.

What open libraries are good for

Community quant libraries now cover screening, portfolio math, and multi-provider data failover. Use them as engines, not as oracles. Read the license. Pin versions. Prefer projects that disclose how they handle missing bars and corporate actions.

Okama-style portfolio tools, multi-provider indicator kits, and simple Streamlit screeners all belong in the surplus catalog when they are free or open and honestly scoped. They do not replace judgment about whether a signal deserves capital.

From notebook to behavior

A backtest that never meets a paper account is unfinished. Take the top three names or setups your stack emits this week and trade them with realistic size rules in Stock Picks. Journal slippage in your own clicks: late entries, moved stops, abandoned plans. The open-source stack measures historical hypotheticals. The paper account measures you.

Standing order

Keep the kit at $0 until the process is clean. Spec every part: license, update cadence, known blind spots. Reject ideas that only work without costs. Graduate to paid data when free rate limits — not ego — become the bottleneck.

This is a field manual entry, not a promise that any strategy profits. Surplus tools remove excuses. They do not print money.