Trading Strategy Development Process

At a glance

Stages: Hypothesis, prototype, validate, forward test, deploy, monitor
Expected survival rate: A small fraction of ideas reach live trading
Key discipline: Pre-registering criteria before testing
Most common error: Promoting a strategy because you are tired of rejecting them

Key takeaways

A research pipeline exists to reject ideas efficiently. If most of your hypotheses survive, your evidence standard is too low.
Write the pass criteria for each stage before running the test, so the decision is not made by looking at the result.
Kill ideas early and cheaply. A quick sanity test that eliminates an idea in an hour is worth more than a thorough test that takes a week.
Keep a research log recording every hypothesis tested, including failures, because that count determines how much evidence a success requires.
Deployment is a stage with its own criteria, not the automatic consequence of a good backtest.

The stages and their gates

Stage	Activity	Pass criterion	Typical survival
1. Hypothesis	State the effect and its mechanism	You can name who pays and why	Most ideas fail here
2. Sanity test	Quick, rough test on a subset	Effect appears in the expected direction	A minority survive
3. Prototype	Full rules, costs, realistic fills	Positive expectancy after costs	Fewer still
4. Robustness	Parameter sweep, multiple markets, regimes	Broad plateau, effect appears elsewhere	A small fraction
5. Validation	Out-of-sample and walk-forward	Efficiency above threshold, no collapse	Fewer again
6. Forward test	Live at minimum size, 30 to 50 trades	Slippage and signal count as expected	Most survivors pass
7. Deployment	Scale gradually on rule adherence	Clean operations, results within distribution	Ongoing

Each stage has a pass criterion defined before the work begins.

Stage 1: generating hypotheses that are worth testing

A hypothesis worth testing names a market behaviour, a reason it exists, and a prediction that could be wrong. Ideas that fail any of those three are not ready for data.

Start from participant behaviour. Who is forced to trade, who is constrained, and who is slow? These produce the most durable hypotheses.
Start from structure. Scheduled flows, expiries, rebalances, and settlement conventions create predictable behaviour that is not a forecast.
Start from documented effects. Momentum, trend, reversal, and carry have extensive literature. Replicating them teaches the workflow before you attempt discovery.
Do not start from a chart. An observed pattern is a result of searching, which means it carries no prior support and requires far stronger evidence.
Write the prediction down. Direction, horizon, and expected magnitude, before testing. This prevents the result from reshaping the hypothesis.
Write the disqualifying condition. What result would make you abandon this? If nothing would, it is not a hypothesis.

The research log

The single practice that most improves research quality is recording every hypothesis tested, including the ones abandoned after ten minutes. The count matters because it determines how surprising a good result actually is.

ID: 2026-047
Date: 2026-04-18
Hypothesis: Stocks gapping down >3% on no news revert
            within 3 days, because the move is liquidity
            driven rather than informational.
Prediction: Positive 3-day return, effect stronger in
            large caps than small caps.
Disqualifier: Negative expectancy after costs, or effect
            absent in large caps.
Universe: US large caps, 2005-2018, point-in-time.
Variants tested: 6 (gap thresholds 2/3/4%, holds 2/3/5 days)
Result: Positive but small; disappears after costs in
        anything below $50m ADV. Survives in top decile
        liquidity only.
Decision: PARK. Revisit if execution costs improve.
Time spent: 3 hours

A minimal research log entry.

Killing ideas efficiently

1
Test the crudest version first
No optimisation, no filters, default parameters. If the raw effect is absent, refinements will only be fitting noise.
2
Apply costs immediately
Many effects are real and unprofitable. Adding costs at the first test rather than the last saves substantial time.
3
Check the direction before the magnitude
If the effect points the wrong way, stop. Do not invert the hypothesis and continue, which is fitting.
4
Test on a second market early
A behavioural effect should appear elsewhere. An effect confined to one instrument is usually an artefact and can be eliminated quickly.
5
Set a time budget per idea
Four hours to a go or no-go decision at the prototype stage. Ideas that require more than that to show promise rarely repay the investment.
6
Park rather than delete
Record the result. Conditions change, costs fall, and a parked idea with a documented reason for failing is a useful asset.

Promotion criteria, decided in advance

Gate	Criterion	Why this threshold
Prototype to robustness	Expectancy above 0.1R after realistic costs	Below this, live degradation consumes the edge
Robustness to validation	Acceptable across a parameter plateau and at least two markets	Evidence of a mechanism rather than a fit
Validation to forward test	Walk-forward efficiency above 0.5; no regime collapse	Suggests the result transfers forward
Forward test to deployment	Signal count within 20 percent, slippage within assumption, adherence above 95 percent	Operational verification, not profitability
Deployment to scale	25 clean trades per size step	Discipline demonstrated before capital increases
Retirement	Drawdown beyond 1.5x backtest, or outside the Monte Carlo distribution	Distinguishes variance from breakage

Frequently asked questions

How many ideas should survive to live trading?

Very few. Professional research pipelines discard the large majority of hypotheses, and most of the survivors are variations on documented effects rather than original discoveries. A pipeline where most ideas pass is one with an evidence standard too low to filter noise.

How long does it take to develop a strategy?

The research phase for a single idea can be days; the full path from hypothesis to scaled live trading typically takes six months to a year because forward testing alone requires months. Attempts to compress that timeline generally do so by skipping validation, which is where the compression later reappears as losses.

Should I develop multiple strategies at once?

In research, yes, because most will fail and parallel work is more efficient. In deployment, no. Adding a second live strategy before the first is executing cleanly means debugging two systems simultaneously and being unable to attribute results to either.

What if I cannot find any strategy that passes?

That is the normal experience and it is informative. The usual responses are to lower the frequency, since slower strategies face less competition and lower costs, to widen the universe, or to accept a smaller edge with correspondingly smaller position sizes. Lowering the evidence standard is the one response that reliably produces losses.

When should I retire a live strategy?

When live results fall outside the distribution your Monte Carlo simulation produced, or when the drawdown exceeds the threshold you defined before deployment. Both criteria must be written down in advance, because in the moment the decision is made under exactly the conditions where judgement is least reliable.

Test this idea before you trade it

Describe the rules in plain language and AlgoTrader AI turns them into a structured strategy blueprint with a configurable historical backtest, cost assumptions, and exportable code.

Build a backtest

Keep reading

Referenced by

Educational use only. This guide explains how a strategy works. It is not investment advice, not a recommendation, and no result described here is a forecast. Test any approach on historical and out-of-sample data, size positions conservatively, and never risk money you cannot afford to lose.