How to know when a strategy is dead (and kill it)

Most bot builders are good at starting strategies. Almost nobody is good at ending them. Accounts die less often from “no edge ever” and more often from keeping a dead edge warm with hope, parameter tweaks, and “it worked last month.”

---

Is a losing week enough to kill a system?

No.

Variance is real. A positive-EV strategy can lose for days or weeks and still be fine — if the sample is small relative to the edge and the risk is controlled.

A losing week is a signal to measure, not a signal to panic-delete.

---

So when is it dead?

Treat a strategy as dead (or suspended) when several of these stack up — not one alone:

1. Expectancy after costs is ≤ 0 over a sample large enough to mean something for your frequency
2. The market regime that fed it is gone (trend → chop, fee structure change, liquidity collapse, competitor crowd)
3. You cannot explain the edge in one paragraph without waving at indicators
4. Drawdown rules fire repeatedly after you already reduced size
5. Fixes only work in-sample — every “improvement” is a re-fit to recent pain

If you need a story to keep it alive, it is already dead.

---

The “one more tweak” trap

This is the most expensive sentence in bot trading:

“Let me just change the threshold / timeframe / stop / pair list once more.”

Sometimes a tweak is legitimate — a bug, a wrong fee model, a broken feed. Those are engineering fixes.

Most of the time, the tweak is grief. You are negotiating with a market that already moved on. Parameter hunting after a break usually produces a system that looks better on the last three months and worse on the next three.

Rule of thumb: if the core idea needs constant surgery to print, the core idea is weak.

---

Measure after fees or do not measure at all

A strategy that is “almost profitable before fees” is not almost profitable. It is unprofitable.

Always include:

• Trading fees
• Spread
• Slippage (especially on thin books and market orders)
• Funding (if futures)
• Your own infra cost if it is material at your size

If the edge only exists in a fee-free fantasy log, it does not exist.

---

Sample size without pretending you are a quant fund

You do not need a PhD. You do need honesty about how many independent trials you have.

• 12 trades is not a system.
• 120 trades on one pair in one regime is still thin.
• Thousands of trades across venues and regimes is where “EV+” starts to mean something.

If your bot trades rarely, accept that time is part of the sample. Do not promote a weekly system to “proven” after one good month.

---

A simple kill checklist we actually use

Before we keep capital on a bot, we answer:

If two or more rows are red, pause. Size to zero is a valid trade.

---

What to do after you kill it

1. Write one page: what you thought the edge was, what broke, what you measured.
2. Archive the code and the config — do not overwrite history with a “fixed” version.
3. Do not immediately deploy the cousin strategy that is 90% the same idea with new lipstick.
4. Return to plumbing and small experiments until a new idea earns size.

Killing a strategy is not failure. Refusing to kill one is how five years of learning turns into a flat account.

---

Where tools help (and where they do not)

Tools cannot tell you an edge is eternal. They can make the measurement less painful:

• Cleaner public market data across venues
• Fewer rate-limit surprises while you research
• Faster iteration from idea → log → decision

If that is useful later:

• Market Data API — public exchange data through a hosted proxy
• nexrade — multi-exchange Python trading library
• nexrade-cache — local Redis-compatible cache

---

Assume every edge can die. Build the habit of checking whether it already has.

That is how we still trade — and how we still delete bots.