CME Group traded an average of 29.4 million contracts a day in the third quarter of 2026, its highest Q3 on record, with double-digit growth in all six asset classes. A strategy you fitted on quieter years now trades in a busier market, and your backtest describes the market it saw. This backtest checklist gives you seven tests, each with a pass mark, a worked MNQ example and a calculator that tells you how high your Sharpe ratio must be once you count every variation you tried.
Key takeaways
- Count every variation you tested. After 240 tries on 4 years of data, luck alone produces an expected best Sharpe of 1.41.
- For 95% confidence that the edge is real, you need roughly (luck hurdle + 1.645) ÷ √years That works out at 2.24 in the example, well above the 1.60 the strategy shows.
- Run the checks in order. A strategy that fails check 2 (costs) has nothing for checks 3 to 7 to measure.
01What CME's record quarter means for your backtest
CME Group reported its third-quarter 2026 results on 21 October. Average daily volume reached 29.4 million contracts, and non-US volume hit a Q3 high of 9 million. Equity index futures and options averaged 7.71 million contracts a day, up 23% from 6.28 million in the third quarter of 2025.
The busy days kept coming after the quarter closed. On 26 October, after Brazil's election, CME's Brazilian real futures traded 498,498 contracts, more than double the previous record of 191,325 set on 28 September, and about eleven times the year-to-date daily average in notional terms.
A backtest fitted on a calmer stretch has never seen this tape. That says nothing about whether your strategy works. It does mean you should run the checks below again before you trust the equity curve, because fills, spreads and the regime your rules depend on may have changed under it.
02The 7-point backtest checklist
Stop at the first check that fails. Each one assumes the ones before it passed.
| # | Check | Test | Pass mark |
|---|---|---|---|
| 1 | No future data | Re-run on data cut at a date; compare every signal before the cut | Zero signals change; fills at the next bar or later |
| 2 | Costs survive | Double the slippage per side | Net per trade still 2× round-turn cost |
| 3 | Enough trades | t = avg net trade ÷ std dev × √trades | t of 2 or more |
| 4 | Beats luck | Deflated Sharpe ratio for the variations you tried | DSR of 95% or more |
| 5 | Holds on unseen data | Out-of-sample Sharpe ÷ in-sample Sharpe | 50% or more |
| 6 | Plateau, not peak | Average the 8 neighbours of your chosen parameters | 70% or more of the best cell |
| 7 | Not one regime or one trade | Split by volatility; drop the best 1% of trades | Both halves net positive; still profitable |
Pass marks for checks 2, 6 and 7 are working rules of thumb. Check 4 follows Bailey and López de Prado's deflated Sharpe ratio. The 50% bar in check 5 is the rule of thumb TradeStation's walk-forward optimiser states for walk-forward efficiency.
Check 1, the truncation test. Run the backtest on data that ends on, say, 30 June, then on the full history. Every signal up to 30 June must match. If one moves, an indicator is reading bars that did not exist yet: a repainting pivot, a daily high used before the close, or a fill at the same bar's close that generated the signal. Fix that before you look at any profit figure.
03Worked example: an MNQ backtest that looks good and fails check 4
A hypothetical intraday strategy on Micro E-mini Nasdaq-100 futures. MNQ is $2 times the index with a 0.25-point tick, so one tick is worth $0.50. The researcher tried 240 parameter combinations on 4 years of in-sample data and kept the best one.
| Input (hypothetical) | Value |
|---|---|
| Variations tried | 240 |
| In-sample years · trades | 4 · 610 |
| Best in-sample Sharpe (annual) | 1.60 |
| Avg net trade · std dev, 1 contract | $15.20 · $118 |
| Costs in test: commission + fees, slippage | $1.50 + 1 tick per side ($1.00) |
| Out-of-sample Sharpe, next 12 months | 0.70 |
Check 2, costs. Doubling slippage adds another $1.00 per round turn. Net per trade falls from $15.20 to $14.20 against $3.50 of total cost: 4.1 times cost. Pass.
Check 3, trades. t = 15.20 ÷ 118 × √610 = 3.18. Pass. You would need about (2 × 118 ÷ 15.20)² = 241 trades at this average and spread to reach t = 2.
Check 4, luck. With 240 variations, the expected best Sharpe from strategies with no edge at all is 1.41 over 4 years. The deflated Sharpe ratio for 1.60 comes out at 65%: about a two-in-three chance the edge is real. Fail. A t-stat of 3.18 looked convincing because it treated the winner as the only test run.
Check 5, unseen data. 0.70 ÷ 1.60 = 44%. Fail, which fits what check 4 found: picking the best of 240 produced much of that 1.60.
Enter your own numbers. The calculator runs checks 2 to 5 and names the first one that fails.
- Costs at 2× slippage $14.20 · 4.1× cost
- Trade t-stat 3.18
- Luck hurdle Sharpe 1.41 · DSR 65%
- Out-of-sample ÷ in-sample 44%
Pass marks: net trade at double slippage still 2× your round-turn cost; t-stat 2 or more; deflated Sharpe (DSR) 95% or more; out-of-sample Sharpe at least 50% of in-sample. The luck hurdle assumes independent variations, normal daily returns and no real skill. Nothing you enter leaves your browser.
04Backtest overfitting: how many variations can you afford?
The luck hurdle comes from the expected maximum of N draws from a normal distribution, scaled to your sample. Bailey and López de Prado give the expected maximum as (1 − γ) × Φ⁻¹(1 − 1/N) + γ × Φ⁻¹(1 − 1/(N × e)), where γ is the Euler-Mascheroni constant, 0.5772. With no real skill, the standard error of a Sharpe estimate is about 1/√T (Lo, 2002), which turns the hurdle into a simple annual figure:
Sharpe needed (95%) ≈ (E[max of N] + 1.645) ÷ √years
For 240 variations, E[max] = 2.825. On 4 years, the hurdle is 2.825 ÷ 2 = 1.41 and the Sharpe you need is (2.825 + 1.645) ÷ 2 = 2.24. Matching the hurdle gets you a coin flip; the extra 1.645 standard errors buy the 95%.
| Variations tried | 2 years | 4 years | 5 years | 10 years |
|---|---|---|---|---|
| 1 (no search) | 1.16 | 0.82 | 0.74 | 0.52 |
| 10 | 2.28 | 1.61 | 1.44 | 1.02 |
| 50 | 2.77 | 1.96 | 1.75 | 1.24 |
| 100 | 2.95 | 2.09 | 1.87 | 1.32 |
| 240 | 3.16 | 2.23 | 2.00 | 1.41 |
| 1,000 | 3.46 | 2.45 | 2.19 | 1.55 |
Annual Sharpe needed for a deflated Sharpe ratio of about 95%, assuming independent variations and normal returns. Correlated variations count as fewer independent trials, so the table overstates the bar for a grid of near-identical settings and understates it if returns have fat left tails.
Read the table in both directions. If you have 5 years of data and your best result is a 1.5 Sharpe, you can afford about 10 honest variations. If you have already run 1,000, a 2.0 on 5 years is not enough. The cheapest fix can be more history: going from 4 to 10 years on 240 variations drops the bar from 2.23 to 1.41.
05Checks 6 and 7: parameter plateaus and regime splits
Check 6. Lay your two most important parameters out as a grid and look at the cells around the one you picked. A real effect fades step by step as you move away from the best setting, while a fitted one collapses one cell out.
Check 7. Split the in-sample trades into two halves by volatility, for example days above and below the median 20-day ATR, and check that each half nets a profit. Then delete the best 1% of trades (6 of 610 in the example) and re-run the totals. If the profit sits in one regime or a handful of outlier days, the strategy is a bet that those conditions return, and a live account will meet a different regime within a quarter or two. A record-volume quarter like Q3 2026 is the moment to run this split again, with the new months included.
06Build the checks into the research engine
These tests are cheap to run once and easy to forget the tenth time. The useful habit is to make the engine log every variation it runs, so check 4 counts real trials instead of the ones you remember. Project Quant Lab, my backtesting and research engine, runs hundreds of strategy variations on 40,000+ bars and checks the survivors on unseen data, which is the exact situation check 4 exists for. In Project Oracle I take a trader's discretionary idea through rules, backtest and walk-forward before it goes live. If you want a strategy built or an existing backtest reviewed against these seven checks, that is the algorithmic trading work I take on. For sizing the result once it passes, see how to size from your drawdown floor.
FAQFrequently asked questions
What should a backtest checklist include?
Seven tests, run in order: no future data (truncation test), costs that survive doubled slippage, enough trades for a t-stat of 2, a deflated Sharpe ratio of 95% after counting every variation, out-of-sample Sharpe at least half of in-sample, a parameter plateau, and profit in both volatility regimes without the best 1% of trades.
How many trades do you need for a valid backtest?
Enough for the t-stat of your average net trade to reach 2: trades needed = (2 × standard deviation per trade ÷ average net trade) squared. With a $15.20 average and $118 standard deviation, that is about 241 trades. A wider spread or a thinner edge needs more.
What is the deflated Sharpe ratio?
The deflated Sharpe ratio, from Bailey and López de Prado (2014), is the probability that a strategy's Sharpe ratio beats the best result luck would produce across all the variations you tested. After 240 variations on 4 years of data, luck alone gives an expected best annual Sharpe of about 1.41.
What is a good out-of-sample Sharpe compared with in-sample?
A common rule of thumb, stated in TradeStation's walk-forward optimiser documentation for annualised returns, is 50% or more of the in-sample figure. Applied to Sharpe, an in-sample 1.60 needs at least 0.80 on unseen data. Below that, most of the in-sample result likely came from fitting.
Sources
- CME Group: third-quarter 2026 results (Form 8-K, exhibit 99.1), 21 Oct 2026 · retrieved 29 Oct 2026
- CME Group: Brazilian real futures set new daily volume record following election, 27 Oct 2026 · retrieved 29 Oct 2026
- CME Group: Micro E-mini Nasdaq-100 contract specifications · retrieved 29 Oct 2026
- Bailey and López de Prado (2014): The Deflated Sharpe Ratio, Journal of Portfolio Management · retrieved 29 Oct 2026
- Lo (2002): The Statistics of Sharpe Ratios, Financial Analysts Journal · retrieved 29 Oct 2026
- TradeStation Walk-Forward Optimizer: walk-forward efficiency · retrieved 29 Oct 2026
Educational content, not financial advice. The MNQ strategy, its figures and the parameter grids are hypothetical examples, not results from a real strategy or backtest. The $1.50 commission is an assumption; use your broker's. CME figures are from CME Group's own releases of 21 and 27 October 2026. Pass marks for checks 2, 6 and 7 are rules of thumb.