Benchmarks

Looking for the indicator library's numbers?

This page is about Wickra Time Machine. Wickra's own indicator benchmarks — the comparison against TA-Lib, talipp, pandas-ta and the other Rust TA crates — live at wickra.org.

The headline figure is snapshots per second — the rate at which the Time Machine re-folds a recorded multi-symbol universe to reconstruct the full microstructure state at a seek target.

Reproduce with:

cargo bench -p timemachine-bench

Measured throughput

Criterion seek group, single machine (Windows, release build, default parallel feature). Each seek reconstructs the full-universe MarketSnapshot by re-folding every recorded event in the prefix; throughput counts events folded (symbols × events/symbol) over wall-clock:

symbols	events/symbol	events folded	seek latency	fold throughput
10	1,000	10,000	~1.6 ms	~6.1 M events/s
10	10,000	100,000	~16.5 ms	~6.1 M events/s
100	1,000	100,000	~16.5 ms	~6.0 M events/s
100	10,000	1,000,000	~182 ms	~5.5 M events/s

The fold rate holds at ~5.5 M–6.1 M events/second whether the universe is 10 symbols or 100, because each symbol folds independently and the cost is governed by total events in the prefix, not by symbol count. A seek to the head of a one-million-event universe therefore reconstructs the whole market in well under a quarter second; a play sweep amortises this by anchoring, reusing the folded prefix between steps (play_10sym_1000ev_step100: 11 full-universe snapshots in ~150 ms).

Method notes

• Numbers are indicative of relative scaling, not a hardware datasheet — absolute values depend on CPU, event count and the indicator set in the spec.
• The parallel (rayon) and single-threaded (--no-default-features, WASM) paths produce byte-identical snapshots; parallelism affects wall-clock only.
• snapshot_interval bounds backward-seek cost: a seek never re-folds more than one anchor window, so seeking near a recent anchor is far cheaper than the head-of-universe figures above.
• Re-bless these figures when wickra-exchange / wickra-backtest are bumped, since engine changes move the per-event fold cost.

The numbers above are the ones in the repository's BENCHMARKS.md, measured with the commands it names.