Point-in-Time Backtesting: Why Most Crypto Backtests Lie

Point-in-time backtesting only uses data that was knowable at each bar's close. Learn the leaks that inflate crypto backtests and how to close them.

Trigr ResearchUpdated 9 min read
On this page
  1. What is point-in-time backtesting?
  2. Why do most crypto backtests lie?
  3. What does "available at the bar's close" mean in practice?
  4. Fill timing and intrabar events in brief
  5. Which costs does a point-in-time backtest include?
  6. How do you know the edge is not luck?
  7. A point-in-time checklist
  8. How Trigr applies these rules
  9. Next steps

TL;DR: Point-in-time backtesting means every decision in the simulation uses only information that existed at that moment: data published by the bar's close, fills at a price you could actually have traded, and costs you would actually have paid. Most crypto backtests break at least one of those rules, and every break pushes results in the flattering direction. The fix is a set of explicit, boring conventions, plus statistics like the Deflated Sharpe Ratio that correct for how many ideas you tried.

What is point-in-time backtesting?

A backtest replays history bar by bar and asks, at each step, what your strategy would have done. The phrase "point-in-time" adds one constraint: at each step the strategy may only see the world as it looked at that moment.

That sounds obvious, but it rules out a lot of convenient shortcuts. You cannot use a daily candle before the day has closed. You cannot use today's revised macro figure when the first published number was different. You cannot fill an order at the same close price that generated the signal, because by the time that close is final, the price has already moved on.

A point-in-time backtest is not a forecast. It is an honest answer to a narrower question: given the rules, the data that was knowable, and the costs you chose to model, what would have happened? Backtests are not guarantees; perps are leveraged and can lose more than expected.

Why do most crypto backtests lie?

They rarely lie on purpose. They lie because each shortcut looks harmless on its own, and because a pleasant result invites less scrutiny than a bad one. The common failure modes fall into six groups.

  • Look-ahead bias. A feature quietly reads data from the future, such as a higher-timeframe indicator that includes the unfinished candle. We cover concrete cases in look-ahead bias in crypto backtests.
  • Impossible fills. The trade is filled at the signal bar's close, or at the exact high or low of a bar, prices that were not tradeable at the moment of the decision.
  • Optimistic intrabar ordering. When a take-profit and a stop-loss both fall inside one candle, the simulator assumes the take-profit happened first.
  • Missing costs. Fees are included but slippage and funding are not, which matters most for strategies that trade often or hold leveraged positions for days.
  • Revised or backfilled data. Macro series get revised, and some alternative data is published with a lag. Using the final value at the reference date is a subtle form of look-ahead.
  • Selection bias. You test 200 parameter combinations and keep the best one. Its Sharpe ratio is partly skill and partly the maximum of 200 noisy draws.

None of these require bad intent. They require only default settings and a spreadsheet mindset. That is why a good backtester makes the honest behavior the default rather than an advanced option.

Why are crypto perps especially exposed?

Crypto perpetual futures combine several conditions that make each leak more expensive than it would be in, say, a daily stock portfolio:

  • Short timeframes. Strategies on 5-minute to 4-hour bars trade often, so small per-trade errors in fills and slippage add up quickly.
  • Leverage. A cost of 0.1% on notional becomes 0.5% of the capital behind the trade at 5× leverage.
  • Funding. Perps charge or pay funding while you hold, a cost that does not exist in spot backtests and that tends to rise in exactly the trending markets many strategies target.
  • Alternative data. Funding, open interest, liquidations, on-chain and macro data each have their own timestamp conventions, and some are published with a delay or revised later.
  • Fast iteration. Tools that make it easy to test hundreds of variants in an afternoon also make selection bias easy to accumulate without noticing.

The rest of this guide walks through each safeguard in turn.

What does "available at the bar's close" mean in practice?

The core rule is simple: a feature at bar t can use anything known at the close of bar t, and nothing else. The hard part is that different data sources become "known" at different times.

Input Naive (leaky) treatment Point-in-time treatment
Daily indicator on a 1H strategy Uses today's daily value on every hour of today Uses yesterday's completed daily value until today closes
Funding rate Uses the rate stamped at a settlement before it has settled Uses only rates already settled by the bar's close
Macro data (e.g. CPI) Stamped at the reference period, using today's revised value Stamped at the release date, using the first published value
CFTC Commitments of Traders Stamped on Tuesday, the "as of" date Stamped after the Friday release
Cross-asset input (e.g. BTC on an ETH strategy) Aligned by timestamp, sometimes reading the next candle Last completed value carried forward, never read ahead

Two public sources show why this matters. The CFTC publishes its Commitments of Traders report on Friday afternoon using data as of the preceding Tuesday, so a backtest that stamps the data on Tuesday trades on information three days early. The St. Louis Fed's ALFRED archive exists precisely because economic series are revised; it lets you retrieve each vintage that was available on a given date.

On Trigr, every feature is computed only from information available at the bar's close, and coarser or cross-asset inputs are carried forward rather than read ahead. API data series carry the time each value became available, and the backtest joins on availability. FRED series are read as first-release values stamped at the date they were first published, and weekly COT data carries a fixed, conservative publication lag. Missing data is never filled with invented values; if a required input is absent, the signal waits.

Fill timing and intrabar events in brief

A signal computed on a bar's close cannot be executed at that same close. The earliest realistic fill is the next bar's open, which is the convention Trigr uses. In 24/7 crypto markets the next open is usually very close to the prior close, so the price difference per trade is small, but it keeps the causal chain honest and it matters a lot when prices jump between bars. See next-bar-open fills and how fill assumptions change results for worked numbers.

Intrabar events need a rule too. A candle only records open, high, low and close, not the order in which the high and low happened. If a long position's take-profit and stop-loss are both inside the same candle's range, you do not know which was hit first.

Trigr handles this conservatively:

  • When the strategy's timeframe is coarser than 5 minutes, it replays 5-minute candles inside the ambiguous bar to see which level was touched first.
  • If that cannot settle it, or finer data is not available, the stop-loss wins.
  • A stop that the market gaps through fills at the worse open price, as a real stop-market order would. A take-profit gapped through fills at the open as well.

This choice costs some headline performance, and that is the point. An optimistic tie-break can turn a mediocre strategy with tight exits into an apparently excellent one.

Which costs does a point-in-time backtest include?

Costs are where many crypto backtests quietly diverge from reality. Trading fees and Trigr's builder fee are always applied. Slippage and funding are opt-in and off by default, so a first result is clearly labeled gross, before those two costs.

When you enable them, each appears as a line item with a before and after delta:

  • Slippage is a flat number of basis points you choose, applied against you on every entry and exit fill. It is not an order-book, spread, market-impact, latency or volatility model, so pick a value that fits the market and order size you are researching.
  • Funding accrues at each roughly 8-hour settlement while a position is held, signed by side, using the source venue's historical series (Binance USDT-M 8-hour settlements) or a flat per-8h rate you provide. Hyperliquid itself pays funding hourly, so the historical series is an approximation of what a Hyperliquid position would have paid, not a record of it.

Here is a hypothetical example of why it matters. Suppose a long-only strategy makes 400 trades over two years with an average gross edge of +0.20% per trade on notional, after fees. Add 5 bp of slippage per side and a flat 0.01% per 8 hours of funding with an average hold of 16 hours.

Line item (per trade, % of notional) Value
Gross edge after fees +0.20%
Slippage, 5 bp entry + 5 bp exit −0.10%
Funding, 2 settlements × 0.01% −0.02%
Net edge +0.08%

Summed over 400 trades and ignoring compounding, that is roughly +80% of notional gross versus +32% net. Nothing about the signal changed; only the honesty of the accounting did. At higher leverage both numbers scale with it. The full breakdown, including how funding behaves for shorts, is in slippage and funding in perp backtests.

How do you know the edge is not luck?

Even a perfectly point-in-time simulation can be fooled by selection. If you try enough variants, one will look good by chance. Bailey and López de Prado formalized the correction in The Deflated Sharpe Ratio (SSRN 2460551): it raises the bar a Sharpe ratio must clear based on how many trials you ran, how long the sample is, and how non-normal the returns are.

On Trigr, the Deflated Sharpe Ratio and the Probability of Backtest Overfitting (PBO) come from Optimize with ML runs. A standard backtest shows raw Sharpe in its headline and lists DSR and PBO as "Not yet computed" in its "Deep dive" section rather than inventing values, so treat that raw Sharpe as optimistic until an ML run or a paper track backs it up. Both statistics are explained in plain language in the Deflated Sharpe Ratio explained.

One more honest boundary: Studio splits full-history strategy backtests by default, holding out the last 30% as out-of-sample beside the in-sample 70%. That period is unseen only until you act on it. Once you choose or tune a strategy after reading it, it is part of your research data. For a final out-of-sample read, freeze the winner and use the anchored, purged walk-forward results from ML optimization where it applies, or collect paper-trading data going forward.

A point-in-time checklist

Use this list on any backtest, whether it came from Trigr, a notebook, or a screenshot someone posted.

Question What a good answer looks like
When is each feature knowable? At or before the bar's close, including release lags
Where do trades fill? At a price available after the signal, such as the next bar's open
What happens when TP and SL share a bar? A conservative rule, stated explicitly
Which costs are included? Fees always; slippage and funding stated as on or off
How many variants were tried? A trial count, and a Sharpe deflated for it
Is there an untouched holdout? Yes, reserved before research, or an explicit statement that there is not
Is missing data invented? No; gaps delay signals rather than fabricate them

If a result cannot answer these, treat its headline numbers as an upper bound. Our guide on how to read a backtest walks through the rest of the report, from drawdown to the trade log.

How Trigr applies these rules

Trigr is built so the conservative choice is the default rather than a setting you have to remember. In one no-code graph you can combine price indicators with funding, open interest, liquidations and macro data across 43 certified crypto perps on Hyperliquid, plus TradFi markets via HIP-3 for backtesting and paper agents only in the current beta. Every run carries an audit trail that flags which inputs were real and which were simulated, alongside the equity curve, monthly returns and full trade log.

The same backtests are available to an AI assistant over MCP, including optional slippage and funding, so a model drafting strategies for you is held to the same conventions. The backtesting docs summarize the cost, intrabar and data-coverage assumptions.

Next steps

Pick one strategy you already trust, rerun it with slippage and funding enabled, and compare gross to net before changing anything else. When you are ready to test ideas at scale, the pricing page shows how many credits each plan includes.

Frequently asked questions

What is point-in-time backtesting?

Point-in-time backtesting simulates a strategy using only the data that was actually available at each moment in history. Every indicator, data feed and decision at a given bar is computed from information known at that bar's close, never from values published or revised later.

Why do crypto backtests look better than live trading?

Most gaps come from a few repeatable leaks: reading future data, filling trades at prices you could not have traded, optimistic assumptions inside a bar, leaving out slippage and funding, and testing many variants but reporting only the best. Each one inflates results in the same direction.

Does a Trigr backtest have an out-of-sample period?

Yes. By default Trigr holds out the last 30% of the covered period and scores it apart from the in-sample 70%, but it is unseen only until you act on it. A final test is a period you reserve before research and evaluate once with a frozen strategy.

Are Trigr backtests net of slippage and funding by default?

No. Trading and builder fees are always applied, but slippage and funding are opt-in and off by default, so a first result is labeled gross. You enable them with your own flat slippage in basis points and either historical or flat funding.

Put the idea to an honest test.

Describe a strategy in plain English or from your own AI assistant, backtest it on point-in-time data, and forward-test it on paper before any real money is involved.