From Backtest to Live

Each stage exists because it catches a category of failure the previous one structurally could not. Skipping a stage does not save time — it defers the discovery of a problem to a point where it costs money.

Live data, live timing, simulated fills. Run it for one to two weeks minimum — long enough to span a weekend, an expiry if relevant, and at least one unusual session.

1. 1
   Compare signal timing against the backtest

   Run the same date range in the backtest and check that paper produced the same signals at the same bars. A discrepancy here is a data or timing problem, and it is easier to diagnose now than later.
2. 2
   Watch the session boundaries

   The open, the close, and the intraday square-off. These are where configuration errors surface, and they surface once per day so you do not wait long.
3. 3
   Check every rejection

   A rejected order in paper is a rejected order live. Order type unsupported, quantity below lot size, price off the tick grid — all cheap to fix here.
4. 4
   Confirm the trade count

   Materially fewer signals than the backtest projected usually means feed gaps; materially more means a configuration difference.
5. 5
   Read the log daily

   Not the metrics — the individual trades. Two weeks of paper is too small a sample for statistics but exactly the right size for reading every trade.

Ten to twenty-five percent of target size. Enough that fills and slippage are real and measurable, small enough that being wrong about something is affordable.

• Measure slippage properly. Twenty fills compared against the bar closes they should have executed at. This number is the single most valuable output of the stage — feed it back into the backtest and re-run.
• Watch for partial fills. If live position sizes differ from planned ones, every P&L figure scales with them and the risk calculation was based on a size you did not get.
• Compare the exit-reason mix against the backtest. More stops live than in the test means the stop is being hit intra-bar in a way the replay could not see.
• Note your own reaction. This is not sentimentality. If a normal losing streak at 25% size is uncomfortable, it will be intolerable at full size, and the strategy will be abandoned at the worst possible moment.

Increase size only when live results are consistent with paper and with the cost-corrected backtest. Divergence does not improve with size; it usually gets worse, because larger orders consume more of the book.

Things that have nothing to do with the strategy and will still stop it working.

• Broker credentials expire. Many require daily or periodic re-authentication. A strategy cannot place orders through a stale session.
• Margin must be sufficient. Including for the fully-filled state of any ladder, not just the first entry.
• Prefer broker-side stops for positions you cannot afford to leave unattended. A stop the platform manages internally depends on the platform being connected.
• Know what happens on disconnect. Resting broker orders survive; internally managed logic pauses. Design the risk controls around that distinction.
• Editing does not affect a running instance. A deployed strategy runs its deployed configuration until stopped and redeployed.