Tardis.dev is the closest thing crypto has to a default tick data vendor. Ask around any systematic desk and someone will have a subscription, plus a folder of scripts for replaying its order book feeds. Quant Arb's widely shared data sourcing guide weighs the main tick vendors against each other and calls Tardis “simply better.” We think that assessment is fair. We also sell one of the two products on this page, so apply whatever discount to our objectivity you consider appropriate.
The comparison is still worth writing, because the two products sit at different layers of the stack, and teams routinely buy the wrong layer. Tardis sells raw material: tick-by-tick trades and full-depth order book increments, replayable exactly as the exchanges published them. Aperiodic sells the finished series most buyers eventually construct from that material: validated, point-in-time microstructure and order flow metrics, delivered as parquet files you can backtest against the same afternoon.
What Tardis does well
Credit first. Tardis records tick-level trades, quotes, and L2 depth across the major venues, Deribit options included, with archives reaching back to March 2019. The exchange normalisation is careful, the client libraries are solid, and the replay tooling is well engineered. If you are calibrating a market-making model, simulating maker fills, studying queue dynamics, or building an execution engine, you need this class of data. Buying it from a specialist beats running your own collectors for a year first, and Tardis is a good specialist.
Nothing below disputes any of that. The question is what the subscription buys, and what it leaves you holding.
The subscription is the cheapest part
Tardis pricing is configured per order (you choose exchanges and data plans), with a published minimum of $300 and an FAQ that states plainly that they do not offer discounts. Realistic multi-venue configurations cost a good deal more; we will not quote figures that would go stale, and their pricing page is one click away. The point is that with raw tick data, the subscription is rarely where the money goes.
Raw L2 history for a handful of major venues runs to tens of terabytes compressed. Between you and your first backtest stands a pipeline: reconstruct books from increments, align timestamps across venues, decide what a gap or a re-delivered packet means, compute the spreads, imbalances, and flow series you wanted in the first place, then validate the output against something independent. Teams staff this with data engineers and measure it in months, and the recurring bill for storage, compute, and maintenance typically reaches tens of thousands of dollars a year before a single strategy idea has been tested.
The operational fine print deserves a read too. Tardis applies request and data-transfer limits per account, scaled to the subscription, and its own documentation notes that several replay sessions running in parallel can hit those request limits quickly. Each API key works from one source IP address at a time. The full archive is reserved for Business subscriptions; Academic, Solo, and Pro plans see four years back. These are defensible choices for a vendor whose product weighs terabytes. They are also constraints you inherit the moment a backtest cluster, two researchers, and a CI job all want the feed at once.
What Aperiodic sells instead
Aperiodic starts from an observation that took us years of paying for raw data to accept: most research questions never needed the ticks. They needed the series computed from the ticks. Bid-ask spread and top-of-book dynamics. Multi-level order book imbalance and liquidity. Signed flow split by trade size, with a toxicity score. Slippage and impact at fixed notionals. Up-tick and down-tick run structure, intra-bar realized volatility, VWAP and TWAP, funding, basis, open interest. And candles with honest timestamps: our TrueOHLCV series is aggregated from locally timestamped trades recorded at collection points near the exchange, so the 14:32 bar contains what was observable at 14:32 rather than what an exchange stamped after the fact. Bar-level look-ahead is subtle enough that most people meet it only after it has flattered a backtest.
Every series is point-in-time, attributed to the exchange it came from, and passed through a validation pipeline (outlier detection, cross-venue reconciliation, documented lineage) before publication. Delivery is parquet over a REST API, with a Python SDK that turns a dataset into a dataframe in a few lines. Coverage today is Binance futures and OKX perpetuals from 2020, plus Hyperliquid perpetuals, at intervals from one minute to one day on every self-serve tier; institutional plans add 15- and 30-second bars, full history, and delivery into your own infrastructure. Self-serve pricing runs from $35 to $879 a month.
The practical difference is scale. A year of minute-level metrics for a symbol weighs megabytes, so questions that were cluster jobs against raw ticks (does this flow signal decay in fifteen minutes or two hours? does book imbalance lead price on OKX the way it does on Binance?) become a loop over files on a laptop. The byte ratio against raw L2 is three to four orders of magnitude, and backtest wall-clock time follows bytes more closely than anyone likes to admit.
Side by side
| Tardis.dev | Aperiodic | |
|---|---|---|
| What arrives | Raw ticks: trades, quotes, full-depth book increments | Validated microstructure, flow, and derivatives metrics |
| Venues | Major exchanges, including Deribit options | Binance futures, OKX perps, Hyperliquid perps (expanding) |
| History | Archive to March 2019 on Business; 4 years on other plans | From 2020; full history on institutional plans |
| Granularity | Every tick | 15s to 1d bars (sub-minute on institutional) |
| Storage on your side | Tens of terabytes | Gigabytes |
| Pipeline you build | Book reconstruction, aggregation, validation | None: load parquet and go |
| Time to first backtest | Weeks to months | Same day |
| Entry price | $300 minimum order; no discounts (their FAQ) | $35/month self-serve |
| Built for | Execution research, book reconstruction, MM calibration | Signal research, backtests, ML features |
Which one to buy
If your edge depends on fill-level fidelity, buy the raw data. Simulating queue position, reconstructing books, latency-sensitive execution: no aggregate substitutes for the ticks, ours included, and Tardis is the right kind of vendor for that work.
If your work is signals, cross-sectional research, regime analysis, or features for models, the honest accounting says you will spend months rebuilding, from a raw feed, a slightly worse version of a catalog you could rent from day one. We built Aperiodic after being that team ourselves; the longer version of that story is in our FAQ.
Plenty of desks should run both: a raw feed where execution research demands it, aggregates everywhere else, because a researcher-hour goes further against clean parquet than against an L2 replay. Tardis is the microscope, Aperiodic is the telescope, and the expensive mistake this page exists to prevent is buying a microscope to answer telescope questions.