Backtests vs live results

Five kinds of result

State	Uses	Can establish	Cannot establish
Simulation	Constructed scenarios	Rule behavior under specified assumptions	Historical or live market performance
Historical backtest	Recorded past data	Past-data result under the model and cost assumptions	Future performance or actual fills
Paper/testnet	Current or test data, no equivalent production capital	Workflow, integration and decision behavior	Production liquidity, impact and capital risk
Shadow mode	Live market data without orders	Real-time signal and control behavior	Actual execution or realized P&L
Live	Production orders and capital	Observed result for that account, period and size	Future returns or results at larger scale

A dashboard needs a denominator

“40% return” is incomplete without capital, dates, cash flows, costs, benchmark, drawdown and whether the number is annualized. APY can magnify a short observation window into an unstable annual figure. The calculation method must remain close to the number.

Backtest quality checklist

Data existed and was available at the decision timestamp
Asset universe includes delisted and failed assets where relevant
Parameters were not selected using the reported test interval
Fees, spreads, slippage, funding and borrow are included
Order size respects historical market depth
Drawdown and adverse regimes are visible
Results are compared with a simple baseline

Live-result quality checklist

Account and observation period are defined
Deposits, withdrawals and strategy changes are visible
Returns are net or gross with costs stated
Open positions and unrealized P&L are handled consistently
Maximum drawdown and exposure accompany return
Verification scope and independent access are explained

The same instructions, three execution models

All three examples below are simulations with invented prices and events. None is a historical backtest or a live account. Each starts with $1000.00, requests a purchase of five units, and later requests the sale of the acquired position. Fees are 0.1% of each filled notional. At 2026-09-08T12:02:00Z, any remaining unit is valued at $104.00. There are no deposits, withdrawals, borrow costs or leverage.

Reference-price fills

Time	Action	Requested units	Filled units	Fill price
12:00:00 UTC	buy	5	5	$100.00
12:01:00 UTC	sell	5	5	$104.00

Delayed-price fills

Time	Action	Requested units	Filled units	Fill price
12:00:01 UTC	buy	5	5	$101.00
12:01:01 UTC	sell	5	5	$103.00

Constrained partial fills

Time	Action	Requested units	Filled units	Fill price
12:00:01 UTC	buy	5	3	$101.00
12:00:02 UTC	cancel	2	0	—
12:01:01 UTC	sell	3	2	$103.00
12:01:02 UTC	cancel	1	0	—

Model	Cash	Inventory at final mark	Fees paid	Equity	Net change
Reference-price fills	$1018.98	0 units / $0.00	$1.02	$1018.98	$18.98
Delayed-price fills	$1008.98	0 units / $0.00	$1.02	$1008.98	$8.98
Constrained partial fills	$902.49	1 units / $104.00	$0.51	$1006.49	$6.49

Equity is cash plus marked inventory. Net change is equity minus the original $1,000, with fees already deducted from cash. Do not subtract execution shortfall again: the changed fill prices already capture it here. Calculations retain thousandths of a dollar; displayed totals round to cents. The partial-fill model therefore ends with $902.491 cash plus $104 inventory, not a fully realized $6.491 return. The remaining unit still carries price risk and would incur another fee if sold.

Cancelling an unfilled remainder is not a sale. The example makes cancellation explicit so an outstanding order cannot silently fill after the valuation time. Real order-state behavior must be checked against the venue: Coinbase documents its order events and remaining quantities. The system recovery guide explains why local state alone is insufficient. Short-horizon models such as AI scalping can be particularly sensitive to the difference between assumed and executable fills.

What would make this live evidence?

A real comparison needs the account and observation period, timestamped order and fill records, fee records, cash flows, open positions and a consistent final mark. It also needs the strategy version and inputs available when each decision was made. A deposit increases account value without being trading profit. A screenshot of ending equity cannot establish that record by itself. These tables illustrate the calculation only; they do not supply those missing records.

How to read a performance page

Preserve the evidence label attached to every number. A backtest remains a backtest when it appears on a polished product page, and a live result remains limited to its named account, dates and capital. If the status changes, the explanation should change with it instead of carrying an old interpretation forward.