From Review to Forecasting
Execution Precision
8 min read
autocorrelationequityRSquared
Build forward-looking trading plans from past data, transitioning from reactive review to proactive strategy planning.
Loading
8 min read
Build forward-looking trading plans from past data, transitioning from reactive review to proactive strategy planning.
Real traders don’t just look back at trades. They project forward with sharper clarity, based on what their own edge is telling them.
Read this first. This lesson uses forecasting in one specific sense only: projecting your own future performance metrics — winrate, expectancy, drawdown, losing-streak length — from your historical trade log. It does not teach market prediction. If you are looking for a way to predict next week's price, this lesson will not help you, and nothing in trading reliably will.
Forecasting in trading is the statistical projection of your own future performance distribution from your historical trade log. Given a 200-trade sample at 53% winrate, you can forecast that 8-trade losing streaks will occur roughly once per quarter. You cannot forecast next Tuesday's BTC range. Confuse these two and the rest of this lesson will hurt you.
| Type | Input | Output | Falsifiable? | Useful for sizing? |
|---|---|---|---|---|
| Market forecast | Price/structure | Next price | Rarely | No |
| Self forecast | Your trade log | Your next 100-trade R-band | Yes (compare actual to forecast) | Yes |
| Setup ranking | Per-setup expectancy | Hypotheses to test | After more N | Indirectly |
Prerequisites. You need a working trade feedback loop before you can forecast — without consistent review you have no clean sample to project from. The other anchor is base-rate literacy from the probability module; forecasting without it collapses into prediction.
This is not market forecasting. You cannot predict price. The Efficient Market Hypothesis is wrong in detail but right in spirit for retail timeframes: your edge does not come from knowing where price goes. It comes from knowing your own statistical signature — winrate, expectancy, drawdown distribution — and sizing accordingly. Confusing these two is the most expensive mistake in retail.
A practical consequence: every "forecast" in this lesson carries a sample size and a confidence interval. A claim without those is not a forecast — it is a guess.
Look at your last 50–200 trades — but treat anything below N=50 as a hypothesis, not a forecast. With N=25 the 95% confidence interval on winrate is roughly ±20 percentage points. You are not forecasting yet; you are generating candidates to test on the next 100 trades.
For each setup, compute the Wilson 95% confidence interval on its winrate, not the naive ratio. The Wilson interval is the standard binomial-proportion CI and avoids the boundary problems of the naive Wald interval at small N. Treat the lower bound as your honest estimate.
Expectancy per trade = (winrate × average win in R) − (loss rate × average loss in R). This is the only number that tells you whether the setup makes money on average. Setup ranking by raw winrate is the textbook small-sample error.
Wilson 95% CI half-width on winrate shrinks with sample size
Diminishing returns: doubling N from 100 to 200 only narrows the band from 10pp to 7pp.
These questions surface which setups have produced positive expectancy in your sample — not where you will win next. With 20–30 trades the confidence interval on any single setup's winrate spans ±15–20pp; treat rankings as hypotheses, not forecasts.
| Insight | What it actually tells you |
|---|---|
| What setup had the best R-multiple? | Candidate for more reps, but only if N ≥ 50 |
| What setup caused the most frustration? | Behavioral signal — drop or reduce size, regardless of WR |
| Where did emotion most interfere? | Identify the trigger context; do not treat it as a forecast input |
| What structure repeats most often? | Sample-size pump — gives you faster CI convergence |
The bias map is a plan, not a forecast. It tells you which hypotheses you are willing to risk capital on for the next 10–20 trades, and at what size. It is grounded in Step 1 numbers.
| Area | Example | Numeric anchor |
|---|---|---|
| High-conviction setup | OB + BOS + imbalance confluence | N=72, WR 58% (CI 47–69%), expectancy +0.42R |
| Setup under review | Sweep without BOS | N=18, WR 39% — too small to act, hypothesis only |
| Avoid for now | Fade against HTF structure | N=24, expectancy −0.15R |
| Mental focus | Patience at open, no impulse trades | Plan-follow target ≥ 80% |
| Performance target | Plan-follow rate, not R-multiple | ≥ 80% over next 20 trades |
This becomes your trading compass for the next 10–20 trades. Each row is a falsifiable claim.
A forecast card is a setup spec card plus base rates. Without sample size and confidence interval, this is a setup spec — useful, but not a forecast. Add the fields a forecast actually requires.
HTF bias bullish (higher-low above support). Invalidation: no BOS after sweep means no entry. Sample drawn from Q4-2025 medium-volatility regime; re-validate if ATR percentile shifts more than 1 sigma.
N=47 trades, historical winrate 26/47 = 55.3% (Wilson 95% CI 41 to 69 percent). Average win +1.4R, average loss -1.0R. Expected max losing streak (95th percentile) ceil(log(0.05) / log(1 - WR)) approximately 4 trades; longest streak in 100 trades roughly 7 trades.
(WR x avgWin) minus (LR x avgLoss), measured on the N=47 sample.
E[R] = 0.553 * 1.4 + 0.447 * -1.0 = 0.30R95th percentile streak length at WR 55.3 percent. Plan-grade, not worst-case.
ceil(log(0.05) / log(1 - 0.553)) approximately 4Plan for at least one streak this long over the next 100 trades. If your sizing cannot accept it, the sizing is wrong, not the setup.
Expectancy x trade count. The 1 sigma envelope is the band you should expect to live inside; outside the band, regime change is the working hypothesis.
| Field | Spec card (insufficient) | Forecast card (this lesson) |
|---|---|---|
| Entry / stop / partial logic | Yes | Yes |
| Sample size N | No | Required |
| Historical WR with 95% CI | No | Required |
| Expectancy in R | No | Required |
| Expected max losing streak | No | Required |
| Regime tag | No | Required |
The scorecard becomes your repeatable forward-play instrument — not just a memory.
Each Sunday or Monday, run a checklist that produces numbers:
"I will take only 1 OB-reclaim setup per session. I will pre-plan entry/exit levels. I will not chase price outside my zone."
The ritual is not motivation — it is calibration. The forecast bands here become the expected envelope for equity curve analysis; when actual equity breaches the forecast band, regime change is the working hypothesis.
At least 50 for a directional hypothesis, 100 for sizing-grade decisions, and 200+ for forecast-grade claims with a usable confidence interval. Below N=50 the Wilson 95% CI on winrate is wide enough (±14pp or worse) that ranking setups against each other is statistically meaningless.
No. You can rank setups by historical expectancy and size positions proportional to confidence — that is not the same as predicting they will continue to work. Setups decay when the regime that supported them shifts; the forecast is a null hypothesis to be falsified by next month's data, not a guarantee.
Recency bias plus small-sample noise. With N=20, even a 70% observed winrate has a Wilson lower bound near 46%. Twenty consecutive winners is consistent with a true 50% winrate over a long enough run. Do not increase size on the basis of a 20-trade stretch.
Sample size N, historical winrate with 95% confidence interval, expectancy in R, expected maximum losing streak, and a regime tag. A spec card without these is a plan — useful, but it cannot tell you whether the setup is worth the size you are putting on it.
Your trade log is a sample from a distribution. The job is to estimate that distribution well enough to size correctly and stop early when it shifts. That is the only forecast worth making — and it has nothing to do with predicting price.
Use the past to bound expectation, not to predict it. The gap between forecast R and realized R is then decomposed in the next lesson via MAE/MFE slippage.
Lesson 5 of 11 — Measuring Slippage with MAE/MFE: now that you can forecast your own R-distribution, the next step is to attribute the gap between forecast and reality to slippage and execution drag.