More granular market data can make a strategy backtest less trustworthy. Tick data gives you more rows, but it also gives you more ways to mistake measurement detail for knowledge about what you could have traded.
That sounds backwards if you have ever watched a bar-based test miss a stop, a spread widening, or a brief touch of your limit price. Those are real problems. The question is whether the finer data actually resolves them, or just lets the simulator tell a more confident story.
What does tick data let a backtest see?
A one-minute bar compresses a sequence of trades into four prices and a volume total. From the bar alone, you usually cannot tell whether the high came before the low, how long price stayed near either extreme, or whether a particular price traded enough size to support your order.
Trade ticks restore the sequence of reported trades. Quotes add the displayed bid and ask through time. With both, you can ask more precise questions about spread, quote changes, and whether a marketable order would have crossed the book at a plausible price.
But a trade print is not a queue position. If your resting order would have joined a level with 18 BTC already ahead of it, seeing 0.3 BTC trade there does not prove your order filled. The tape tells you what happened in the market; it rarely tells you exactly where your hypothetical order sat.
Where can finer data make the result worse?
The first risk is uneven coverage. Tick archives often have missing intervals, venue gaps, symbol changes, or different retention windows across markets. A strategy tested only where the vendor has complete history can quietly become a strategy for the easiest-to-measure assets and periods.
The second risk is timing. Timestamps may represent exchange event time, vendor receipt time, or the moment a file was assembled. If your simulator sorts events from different feeds as though their clocks were perfectly synchronized, it can react to a quote before that quote was available to your process.
And more rows make implementation errors easier to hide. A bar strategy may make one decision per minute. A tick-driven version can trade dozens of times in that minute, each time paying spread, fees, and impact. If the simulator assumes every signal can execute at the next observed price, its precision is decorative.
The 60 observations are not automatically 60 independent chances to trade. They may all describe the same evolving quote, and your order can affect what fills remain available. A high-frequency simulation needs an order and latency model that uses the extra detail honestly.
When is tick data worth the cost?
Start with the question you cannot answer using your current data. If it is “which came first, my stop or my target?”, trade ticks may help. If it is “how much did the spread cost at entry?”, quote data is more relevant. If it is “would my passive order have filled?”, you may need depth and queue assumptions, and even those rarely recover the full historical queue.
| Research question | Useful data | What it still cannot prove |
|---|---|---|
| Did a threshold trade before another threshold? | Trade ticks | That your order filled at either threshold |
| What spread did a market order cross? | Timestamped bid and ask quotes | That displayed size remained available |
| Could a passive order fill? | Quotes, depth, and queue model | Your exact place in the live queue |
For a daily signal held for weeks, one-minute bars may already be finer than the decision process needs. For a strategy whose edge depends on execution over seconds, bars can erase the very effect under study. Match data resolution to the strategy’s decision and execution horizon.
How do I tell whether the added detail helped?
Keep the strategy and test period fixed. Compare the bar-based and finer-data versions on the same decisions wherever possible. Track changes in fills, turnover, spread paid, fees, and market impact alongside return and risk statistics. If the Sharpe rises while trade count triples and the execution model still fills at the next print, the improvement deserves suspicion.
Then degrade the fine data deliberately. Drop a small random share of events, add realistic feed delay, and widen the assumed spread or impact. A result that collapses under a few milliseconds of delay may be a property of the simulator’s event ordering, not a durable feature of the strategy.
One afternoon I once spent tracing a spectacular improvement to a feed that had duplicated a short burst of prints. The model had learned to trade the duplicate. More data had made the bug look like an edge.
Critics of bars are right that aggregation can conceal path-dependent execution. A stop and target inside the same bar cannot be ordered from OHLC alone, and spread snapshots can miss brief dislocations. Tick and quote data can answer those questions more faithfully when the feed is complete and the simulator respects event timing.
The practical standard is simple: buy more detail only when it changes a decision you can describe, and show which assumptions let that detail become a fill. Otherwise, you have a larger dataset and a more precise-looking backtest, with no better evidence that the strategy could trade.
← All posts


