Quant
The deployment ladder — paper, shadow, small, scale
You've validated the strategy. The edge survived walk-forward, the parameters sit on a plateau, the trade count is respectable. Now comes the part nobody writes about: turning it on. The temptation is a single step — backtest to live, full size, today. The discipline is four rungs, and the thing that gets you promoted between them is not profit.
Why not just turn it on
Because your backtest and your live system are two different programs that both claim to implement the same strategy. One reads a file in a loop; the other handles a socket that drops, a broker that rejects, a clock in the wrong timezone, and a position that's 40% filled. They will disagree. The only question is whether you find out on a rung, or with real size on.
What each rung actually catches
- 1 · Paper. Your strategy runs against a live data feed and pretends to trade. This catches the embarrassing class of bugs: crashes at midnight, timezone and daylight-saving errors, feed gaps, holidays, orders your broker would reject outright. What it emphatically does not catch is fills — paper fills are as fictional as backtest fills, and for exactly the same reasons.
- 2 · Shadow. The strategy runs live, produces real decisions with real timestamps, and places nothing. You record what it would have done, then feed the identical period through your backtest engine and compare. This is the highest-value rung and the most skipped: it's the only place you catch your two programs disagreeing while it's still free.
- 3 · Small live. Real money at the minimum size the market allows. Now — and only now — you learn about fills, latency, partial fills, fees, and rejects. The P&L is irrelevant at this size. The divergence is the entire point.
- 4 · Scale. Increase size gradually, watching for the thing only size reveals: your own market impact. An edge that works at 100 shares can vanish at 100,000 because you've become the thing moving the price against yourself. Capacity is discovered, never assumed.
Reconciliation is the gate
At every rung you're asking one question: given the same inputs, did live do what the backtest said it would? Same bars in, same signals out, same intended orders, same times. Any unexplained difference is a bug until proven otherwise — and "unexplained" is doing real work in that sentence. A divergence you can explain ("we filled 2 ticks worse; that's spread") is information. A divergence you can't is a defect you don't understand yet, and it will not stay small.
This makes the promotion rule uncomfortable but simple: never promote a rung because it made money. Promote it because it did exactly what you predicted it would do. A rung that's up 3% while diverging from expectation is more dangerous than one that's flat and tracking perfectly — the first is a broken system currently being paid by luck, and luck reverses.
Don't flip a switch from backtest to real money. Climb: fake money on live data, then real decisions with no orders, then tiny real orders, then size. At each step check it did what you predicted — not whether it profited.
Going back down
The ladder runs both ways, and this is what makes it a discipline rather than a checklist. You demote when tracking error grows without explanation, when the live trade distribution drifts from the backtest's, or when you change anything meaningful in the strategy — a change is a new strategy and it starts at a lower rung, however small the diff looked.
What makes all of this possible is the boring infrastructure underneath: an execution engine that records exactly what it intended and what actually happened, hard limits that make a bug on rung 3 cost lunch money instead of the account, and enough observability to answer "why did these two disagree?" months later. Get on the ladder before you have an edge to deploy — the climb is far more pleasant when you're not impatient to earn.
And once you're at full size, the ladder's central question doesn't retire. It just becomes the tripwire: a live strategy that stops matching its own expected behaviour is telling you something, and knowing when to turn it off is the last rung of all.