How to Build a Trading Strategy from Scratch: Without Curve-Fitting

Most "build your own strategy" guides start with indicators and chart patterns. That's backwards. A strategy starts with a market hypothesis, not a setup, and the difference is why 90% of homemade strategies look great in backtest and lose money live.

This is the 7-step framework that produces strategies that actually survive live trading.

Step 1: Pick a market hypothesis (not an indicator)

A strategy needs a reason to work. The reason is a hypothesis about how a specific market behaves at a specific time.

Bad hypothesis (indicator-led):

• "RSI is oversold so price goes up"

Good hypothesis (market behaviour-led):

• "During the first 90 minutes of NY cash open, the S&P pulls back to the 21 EMA on the 5-min in roughly 65% of trending days, because institutional algos buy dips on confirmed trends"

The second one is testable, observable, and has a structural reason to exist. The first one is just a math relationship that may or may not have edge.

Write your hypothesis in one sentence:

• WHO is causing the behaviour (institutions, retail, algos, hedge flows)
• WHEN it happens (specific session, day, time)
• WHAT the pattern looks like
• WHY it exists (the structural reason)

If you can't fill in all 4, your hypothesis isn't good enough.

Step 2: Define the setup precisely

Now turn the hypothesis into an objective checklist:

Setup: NY Open Pullback (NAS100)
Entry conditions:
  [ ] Time between 14:30-16:00 UK
  [ ] 5-min trend confirmed (3 consecutive higher highs OR 3 consecutive lower lows)
  [ ] Price pulls back to 21 EMA on 5-min, touches within 5 points
  [ ] Rejection candle at EMA: hammer or shooting star, wick > body
  [ ] DXY direction confirms (e.g. long NAS = DXY weak or flat, not surging)
Exit conditions:
  [ ] Stop: 1.5× ATR(14) beyond rejection wick
  [ ] Target: 2R OR opposite swing high/low, whichever is closer
  [ ] Time-stop: close at 17:00 UK regardless

If the rules are subjective ("clear rejection," "good price action") then the strategy can't be tested. Subjective rules will appear to have edge in your manual backtest because you're cherry-picking unconsciously.

Step 3: Manually test 50 trades on past charts

Before automating anything, scroll back through historical charts and find 50 instances of the setup, without looking ahead.

The discipline: scroll to the day, mark the entry, then advance bar-by-bar to see what happens. Don't peek at the right edge before marking entry.

Track for each trade:

• Date / time / instrument
• Entry price, stop price, target price
• Outcome: TP hit, SL hit, time-stop, or partial
• R-multiple: how many R you made (or lost)

After 50 trades, calculate:

• Win rate (% of trades that hit TP)
• Average winner in R
• Average loser in R (usually -1.0 if you respect stops)
• Expectancy = (win rate × avg winner) - (loss rate × avg loser)

Expectancy above +0.3R per trade = solid edge. Expectancy 0 to +0.3R = marginal, needs refinement. Expectancy below 0 = the setup doesn't work, scrap it.

Step 4: Run an out-of-sample test

The next 50 trades should come from a different time period than your first 50. Test 1 = 2022 data. Test 2 = 2023 data. Don't tweak rules between tests.

If your expectancy in Test 2 is similar (within 30%) of Test 1, you have a real strategy. If it drops to half or goes negative, you over-fit to Test 1 data.

This is the single most-skipped step. Most homemade strategies look great because they were tuned on the same data used to evaluate them. Out-of-sample testing exposes curve-fitting.

Step 5: Forward test on a demo account for 30 days

Live demo trading reveals problems backtests miss:

• Spreads are wider in some sessions than your historical data shows
• Slippage on stops in fast markets eats expectancy
• Execution lag can change a "5-point stop" into a "7-point stop"
• Emotional bias, you'll be tempted to skip setups that look "ugly"

Run the strategy for 30 days on demo. Use real position sizes (your planned account × planned risk %). Take every signal that fires, even bad-looking ones.

After 30 days, compare expectancy to your manual backtest. Expect a 20-30% drop. If it's worse than that, something in the live conditions is killing the edge.

Step 6: Quantify position sizing BEFORE going live

Risk per trade = 0.5-1% of account.

Account: $50,000
Risk per trade: 1% = $500
Setup A average stop distance: 15 points on NAS100
Per-lot risk at 15pt: $15 (assuming $1/pt CFD)
Position size: $500 / $15 = ~33 lots

If your strategy backtest had a 15% max drawdown over 100 trades, your live max drawdown could easily be 25%. Plan for it:

• If you'd be account-busted at 10% drawdown on your prop firm, your strategy risk per trade must be HALF what you'd theoretically need
• 0.5% risk × 100 trades × 15% drawdown = ~7.5% account drawdown maximum

Drawdown math is more important than expectancy math. Strategies that have edge but blow accounts during normal drawdowns aren't tradeable.

Step 7: Journal every live trade

Once live, every trade goes into a journal that tracks:

• Did I follow the setup checklist 100%?
• Was the entry within 1 minute of signal fire?
• Did I move the stop or take the trade off early?
• What was my emotional state at entry?
• What was the actual R outcome?

After 50 live trades, review per-setup expectancy + per-deviation expectancy. You'll find that:

• Setups followed 100% have one expectancy
• Setups where you "improved" the entry have a different (usually worse) expectancy
• Setups taken when you were tired / stressed have a much worse expectancy

The journal data IS the iteration. Without it, you're guessing why the strategy isn't performing.

Common curve-fitting traps

1. Over-fitting to recent data, using only the last 6 months of charts means your strategy is tuned to a single regime
2. Cherry-picking instruments, testing on only winning pairs and ignoring the rest
3. Optimising parameters too tightly, best EMA = 23, best stop = 1.6× ATR, etc. = curve-fit
4. Ignoring spread / slippage, backtest with 0 spread assumption, live trades pay 2-5 pips spread
5. Subjective "I would have skipped that one", every signal that fires in the rule set MUST be counted

If your strategy needs to be perfect to work, it's curve-fit. A real strategy survives missed trades, slippage, and a few subjective skips and still produces edge.

The realistic timeline

• Week 1-2: Hypothesis + setup definition
• Week 3-5: 50-trade manual backtest
• Week 6-7: 50-trade out-of-sample backtest
• Week 8-12: 30 days demo forward test
• Week 13+: Live trading at 0.25% risk for first 30 trades to verify

Total: 3 months from hypothesis to live trading. Strategies built faster than this are usually curve-fit and die within their first month live.

How to track strategy development cleanly

The hardest part isn't running tests, it's organising data across 50 manual trades + 50 OOS trades + 30 demo days + live trades without losing track.

RB Trading Pro Journal has a strategy development module specifically for this: backtest replay, R-tagged setups, OOS analytics, and live-vs-test expectancy comparison. Risk-free for 30 days.

TL;DR

7-step framework:

1. Market hypothesis (not indicator)
2. Setup as objective checklist
3. 50-trade manual backtest with expectancy
4. 50-trade out-of-sample to detect curve-fit
5. 30-day demo forward test
6. Quantify position sizing for max drawdown
7. Journal every live trade, iterate from data

Strategy building is slow because it has to be. The traders who skip steps are the ones whose perfect-looking backtest goes to zero in month 2 live.