Backtest vs Live Results: Why They Differ and What to Trust

What a backtest really is

A backtest is a simulation. The platform replays historical price data through the logic of a strategy and records the trades it would have taken. Nothing more. It is not a recording of real trades, it involves no real broker, no real latency, and no real money changing hands. The output is a hypothesis about how the strategy would have behaved, conditional on a long list of assumptions baked into the test.

The quality of that hypothesis depends entirely on two things: the quality of the historical data and the honesty of the assumptions. Feed a backtest smoothed one-minute bars with a fixed spread, and it will happily report results for a scalping strategy that could never have been achieved in the real market. The simulation does not know it is lying, and neither does the person running it — unless they go looking.

Where the gap comes from

When a live account underperforms its own backtest, the difference is rarely one big cause. It is usually a stack of small, compounding frictions, each of which the simulation either ignored or modeled optimistically.

Spread widening and slippage. Most historical data assumes a fixed, average spread. Real spreads breathe: they widen around news releases, at rollover, during the Asian session on some pairs, and in any moment of thin liquidity. Slippage works the same way — a backtest fills every order at the exact requested price, while a live server fills orders at whatever price is actually available. For strategies that target small moves, these two frictions alone can consume a large share of the theoretical edge.

Execution latency and requotes. In simulation, orders execute instantly. In reality, an order travels from the terminal to the broker server, gets processed, and comes back. That round trip takes time, and during fast markets the price may have moved before the order lands. Some brokers respond with requotes; others simply fill at the new price. Either way, the entry the backtest assumed is not the entry the live account received.

Tick data quality and modeling mode. MetaTrader can run tests in several modes, and they are not equal. A test built on interpolated one-minute bars invents the path price took inside each bar. A test in Every tick mode is better, and a test run on real tick data with variable spreads is better still. The modeling quality figure the strategy tester prints is a data-coverage metric, not a guarantee: a test can show high modeling quality and still rest on synthetic ticks. For strategies sensitive to intrabar movement — tight stops, tight targets, trailing logic — the modeling mode can change the character of the results entirely.

Broker differences. The same expert advisor on the same pair can produce different results at two brokers, because the inputs differ:

• typical and worst-case spread on the traded symbols
• commission per lot, which some backtests forget to include
• swap rates on positions held overnight or across the weekend
• stop level and freeze level restrictions on pending orders
• server time zone, which shifts where daily candles open and close

Changing market regimes. Perhaps the deepest cause: a strategy is always tuned, deliberately or not, to the conditions it was tested on. A system built during a low-volatility, ranging period meets a trending, high-volatility market and behaves differently — not because anything is broken, but because the market it was shaped by no longer exists. Backtests average across years; live trading happens in whatever regime is running right now.

Overfitting: the silent killer

Everything above explains why a good strategy loses a little in translation. Overfitting explains why a bad strategy can look brilliant in testing and fall apart immediately in live trading.

Overfitting, or curve fitting, happens when a strategy is optimized until it matches the accidents of one specific slice of history. Every added parameter — one more filter, one more session window, one more magic threshold — gives the optimizer another dial for memorizing the past instead of learning something general about the market. The equity curve gets smoother and smoother in the tester, and the strategy gets more and more fragile in reality. The uncomfortable part: the process feels like careful research while it is happening.

The difference between a robust edge and a memorized past is that a robust edge survives contact with data it has never seen. Two techniques probe for that, imperfectly:

• Out-of-sample testing: optimize on one span of history, then evaluate untouched on a later span. If performance collapses outside the optimization window, the tester was measuring memorization.
• Walk-forward analysis: repeat that split in rolling steps across the whole dataset, re-optimizing on each window and testing on the next. It asks whether the process of tuning the strategy — not one lucky parameter set — keeps working over time.

Both are partial defenses, not proofs. A determined optimizer can overfit a walk-forward procedure too, by rerunning it until it looks good. They raise the bar; they do not remove the risk.

Operational reality

Live trading also adds a category of failure the strategy tester cannot represent at all: operations. A backtest never loses power. A real setup does. Mundane examples that show up in live results and nowhere in a simulation:

• a VPS restarting for maintenance while a position is open
• a dropped connection during a news spike, missing an exit
• a terminal update that silently disables expert advisors
• a broker-side symbol change that stops the robot from trading
• human interference — closing a drawdown early, skipping a signal, changing settings mid-trade — which is by far the most common one

None of these are strategy flaws, but all of them shape the live track record. Over a long enough period, operational noise alone guarantees that live results will diverge from any simulation.

So what is a backtest good for?

After all this it might sound like backtests are worthless. They are not — they are simply answering a different question than most people ask of them. A backtest is weak evidence about future returns and reasonably good evidence about strategy character:

• How often does it trade, and at what times?
• How long do losing stretches last, and how deep do they go?
• Does it behave consistently across different years and regimes?
• Does it depend on a few outlier trades, or on many small ones?
• Does the logic survive stress: spread shocks, gaps, missed bars?

Used this way, a backtest is a stress-testing and expectation-setting tool. It tells you what living with the strategy might feel like — the rhythm, the drawdown profile, the patience required — even if it cannot tell you what the account balance will be next year.

How to weigh backtest vs live

A simple hierarchy covers most situations:

• A verified live track record beats a backtest of any length. Months of real execution on a third-party verified account contain information no simulation can produce: real spreads, real slippage, real operations, real regime exposure.
• A young live record plus a long backtest is a reasonable middle ground — if read with care. The live account validates that execution matches the simulation so far; the backtest sketches long-run character. Check that the live period actually resembles the corresponding stretch of the backtest. Broad agreement is a good sign; a live account already diverging from its own simulation within weeks is not.
• A backtest alone is a hypothesis, not evidence. It can justify a demo trial or a small-scale forward test. It cannot, by itself, justify confidence.

A practical checklist

When a vendor shows you both a backtest and a live account, walk through this before anything else:

1. Is the live account verified by a third-party tracking service, with broker and account type visible — real or demo?
2. Was the backtest run on real tick data with variable spread, and is the modeling mode disclosed, or only a bar-based approximation?
3. Are commissions, swaps, and realistic slippage included in the test, or does it assume frictionless fills at a fixed spread?
4. Does the backtest span multiple years and multiple market regimes, or one favorable stretch?
5. Is there any out-of-sample or walk-forward evidence, or was the entire history available to the optimizer?
6. Does the live period overlap the backtest period, and do the two look similar over the overlap — trade count, rhythm, drawdowns?
7. How does the vendor talk about drawdowns and losing months? Open discussion is a good sign; silence is not.
8. Can you run the strategy on a demo account or with minimal risk first, and verify its behavior against both records yourself?

None of these steps require special tools — only the willingness to treat a beautiful equity curve as the beginning of an investigation rather than the end of one.