QUESTION
Where does binary floating-point representation corrupt market data, and does it matter?
- 01
HYPOTHESIS
Float representation of prices and monetary values accumulates representation error that surfaces in aggregates, joins and risk calculations.
- 02
METHOD
Comparison of ingestion-storage-replay paths under exact decimal arithmetic versus IEEE-754 floats, tracking divergence through aggregation.
- 03
EVIDENCE
quant-market-data-replay stores and replays with exact decimal arithmetic across the full pipeline.
- 04
RESULT
Representation drift is eliminated as a failure mode; joins and aggregates behave identically across runs.
- 05
LIMITATION
Exact arithmetic trades throughput for correctness — appropriate at research speeds, costly at HFT scale.
- 06
CONCLUSION
In research infrastructure, correctness of representation is cheap insurance against an invisible failure class.
EVIDENCE SOURCE
QUANT MARKET DATA REPLAY
Every backtest is a claim about the past — and most are wrong in a specific, preventable way: they let strategies see the future.
METHOD — TEST SUITE · DETERMINISM