Three checkpoints in this guide
Follow the full walkthrough in order, or jump directly to one of its main sections.
This is a transparent product walkthrough, not a customer testimonial. The 500 closed trades below form a synthetic demo dataset designed to show how a useful AI Coach review should move from scope to observation to a testable next action. No person named “Sarah” supplied these records, and no live trader outcome is claimed.
The valuable part is still real: a large journal can hide a weak time window, an instrument drag, or a sizing inconsistency that is hard to see trade by trade. TSB can surface and explain those patterns only from a server-owned evidence set. It must preserve the account, date range, eligible trade count, currency or R basis, exclusions, evidence cutoff, and exact membership behind the answer.
Scale note: Trader’s Second Brain has processed 600K+ imported trades cumulatively. That proves operating scale, not that this walkthrough analyzed the entire corpus. Every number below uses the 500-trade demo cohort or a smaller named eligible subset.
What the 500-trade demo contains
The walkthrough fixes the scope before asking the Coach anything. That prevents a later filter change from silently altering the denominator.
| Scope field | Demo value | Why it matters |
|---|---|---|
| Account | Demo Forex / USD | Results from another account are not mixed in. |
| Evidence window | January 1–December 31, 2025 | The answer describes a fixed historical period. |
| Eligible population | 500 normalized closed trades | Open positions and non-trade cash events are outside the denominator. |
| Common fields | Date/time, instrument, side, realized result in R | All 500 trades can enter the time and instrument views. |
| Sizing fields | 482 of 500 complete | Only these trades can support a risk-percent comparison; 18 are excluded from that lens. |
| Evidence cutoff | December 31, 2025 | Later trades do not rewrite the stored First Answer. |
The three analyses below overlap. The time table totals 500 trades, and the instrument table also totals the same 500 trades. They are different lenses over one cohort, so adding their row counts together would double-count evidence. The sizing lens is narrower: 54 subject trades plus 428 controls equals 482 eligible trades, with 18 explicitly excluded.
What the Coach can actually claim
The current TSB server contract separates deterministic metrics from trader-written context. Account scope, period, sample size, metric values, and evidence membership are server-authoritative. Notes such as “I felt impatient” remain trader-provided context; they can help frame a question but cannot become a measured motive.
A Coach answer may cite only numeric observations included in that trusted evidence contract. If the model introduces a number the evidence set does not contain, the server validator rejects it. If there are no observations, the answer becomes insufficient evidence instead of filling the gap with a plausible story.
That boundary makes the Coach more useful, not weaker. It can say “the selected 160 trades averaged less R per trade than the 340-trade control” and link back to the exact rows. It cannot jump from that observation to “fatigue caused the loss” unless fatigue was separately and credibly measured.
Demo finding 1: a time-window drag
The first breakdown compares entries before 14:00 with entries at or after 14:00 in the demo account’s recorded timezone. Every one of the 500 trades belongs to exactly one row.
| Entry window | Eligible n | Net result | Average / trade |
|---|---|---|---|
| Before 14:00 | 340 | +90.0R | +0.265R |
| 14:00 or later | 160 | -30.0R | -0.188R |
| All eligible trades | 500 | +60.0R | +0.120R |
The arithmetic reconciles: 340 + 160 = 500, and +90.0R − 30.0R = +60.0R. The observed difference between subject and control averages is about −0.453R per trade. That is a strong candidate for inspection, but it is still an association inside a synthetic historical cohort.
What to inspect before writing a rule
- Is “14:00” in a consistent timezone, including daylight-saving transitions?
- Does the late window contain a different setup, instrument, or day-of-week mix?
- Are fees and converted P&L complete for both groups, or is R the only comparable basis?
- Did the cutoff get chosen before the result was seen, or was the hour tuned to maximize the apparent gap?
The Coach can direct the reader to those exact 160 trades. A human still decides whether the next step is a temporary restriction, a setup-specific rule, or no change. The broader guide to filtering a trading edge explains why one-variable cuts should be tested before stacking filters.
Demo finding 2: instrument results that reconcile
The same 500 records can be regrouped by instrument. This does not create another 500 observations; it changes the question asked of the same evidence.
| Instrument | Eligible n | Net result | Average / trade |
|---|---|---|---|
| EUR/USD | 210 | +43.0R | +0.205R |
| USD/JPY | 120 | +28.0R | +0.233R |
| GBP/USD | 105 | -6.0R | -0.057R |
| GBP/JPY | 65 | -5.0R | -0.077R |
| All eligible trades | 500 | +60.0R | +0.120R |
Again the totals close: 210 + 120 + 105 + 65 = 500, while +43.0R + 28.0R − 6.0R − 5.0R = +60.0R. It is legitimate to say that the two GBP buckets were negative in this demo window. It is not legitimate to say GBP pairs are structurally bad, that the instruments caused the losses, or that removing them would reproduce the historical total.
There is also overlap with finding 1. Some GBP trades may sit inside the late window. Removing both groups and adding their historical losses would count those trades twice. The exact-membership link is what lets an analyst measure that overlap instead of telling a cleaner story than the data supports.
Demo finding 3: sizing drift with a smaller denominator
The sizing question needs an account-balance snapshot and an intended-risk field. Eighteen trades lack one of those fields, so this lens has 482 eligible records rather than 500.
| Sequence group | Eligible n | Average actual risk | Observed difference |
|---|---|---|---|
| After 3+ consecutive wins | 54 | 1.13% | +0.12 percentage points vs control |
| All other eligible trades | 428 | 1.01% | Control |
| Excluded from this lens | 18 | Not calculated | Missing balance or intended-risk evidence |
The honest finding is narrow: within this demo, actual risk was higher in 54 trades following a three-win sequence than in 428 other trades with complete sizing evidence. The table does not prove overconfidence, and it does not say the remaining 18 trades had normal sizing. Their value is Not calculated, not zero.
This is where a trade-by-trade journal and a Coach complement each other. The system can compute the sequence and surface the difference; the trader can inspect the exact orders and decide whether rounding, contract increments, balance changes, or a deliberate rule explains it. If position size itself is uncertain, use the position-size calculator as an input check rather than treating the Coach label as the calculation.
How one finding becomes a Current Focus
A weak workflow implements all three findings at once and celebrates any later improvement. That destroys attribution. A stronger workflow promotes one observation to a falsifiable rule while freezing the original scope.
- Freeze the First Answer. Store the account, evidence window, 500 analyzed trades, cutoff, selected comparison, and exact membership.
- Inspect the subject rows. For the time finding, review the 160 late-window trades for setup mix, instrument overlap, missing costs, and timezone errors.
- Write one bounded action. In this demo: pause new entries after 14:00 for a short recheck, while leaving setup and risk rules unchanged.
- Define comparable evidence. The product can track progress toward a fixed recheck target; a ten-trade checkpoint is an operational checkpoint, not proof of a durable edge.
- Keep or reverse the rule. Compare the new eligible records with the frozen observation, state the uncertainty, and reopen the exact trades before deciding.
This is the part that makes the Coach more than a chat box: an observation can remain linked to a reviewable evidence set and a later checkpoint. The trade-review workflow shows how to record the context that deterministic fields cannot supply.
How to use 600K+ without breaking the logic
The product-scale statement and the analysis result answer different questions:
- 600K+ imported trades processed describes cumulative ingestion scale across TSB.
- 500 demo trades is the fixed cohort used in the first two walkthrough tables.
- 482 sizing-eligible trades is a smaller cohort with the fields required for the third question.
- 54 post-streak trades is the subject group for one comparison, not “54 of 600K+.”
The same rule applies to real first-party research. If TSB reports 3,000 eligible shorts and 2,000 eligible longs, the direction-classified denominator is 5,000 trades. The rest of the cumulative corpus may belong to another date range, lack a reliable side field, be outside the account scope, or be unavailable under the current retention/query contract. The article must say which.
Why a clean pattern can still be wrong
More rows do not rescue a poorly chosen question. If an analyst tries many hours, instruments, weekdays, setups, and sequence definitions, one attractive result can appear by selection alone. The classic false-discovery literature exists because testing many hypotheses changes the error problem; the Benjamini–Hochberg paper formalized one approach to controlling that risk.
Likewise, a confidence threshold is not a truth label or a measure of practical importance. The American Statistical Association’s statement emphasizes context, study design, and the full body of evidence. For a trading journal, the practical response is simple: show the effect size and exact cohort, disclose how many cuts were explored, avoid tuning a threshold after seeing the answer, and test the proposed rule on later comparable trades.
There is no universal “200 trades means reliable” switch. Required sample size depends on variance, effect size, the decision cost, and the desired precision; NIST’s treatment of confidence limits makes the dependence on sample variability and sample size explicit. Use the sample-size guide for trading decisions to frame that question without pretending one count validates every segment.
What TSB does—and does not do
TSB is our product. Its current observed-trading layer can group eligible records by setup, session, weekday, instrument, and side; report coverage and exclusion reasons; retain bounded evidence references; and give the Coach a deterministic, read-only evidence contract. The Coach can explain those observations and point to a server-owned next action.
TSB does not turn a historical association into causality. It does not infer fatigue, fear, or overconfidence from P&L alone. It does not promise that removing a losing bucket will preserve the rest of the strategy. It also fails closed when the selected evidence is missing instead of fabricating a personalized answer.
To reproduce this walkthrough with your own data, first reconcile the import, choose one account and range, check field coverage, and ask one bounded question. The practical AI Coach guide continues from prompt design through evidence inspection and recheck.
Review one evidence set—not a marketing story
Open TSB Coach from a specific Journal, Dashboard, Leak Map, setup, or replay context. Verify the scope and exact trades before accepting any suggested rule.
Open AI CoachThe five checks before you act
- Scope: account, date range, timezone, currency or R basis, and evidence cutoff are visible.
- Membership: subject, control, exclusions, and overlap can be traced to exact records.
- Math: counts and totals reconcile; percentages name their denominator.
- Meaning: observation, inference, and trader-written context are not blended together.
- Recheck: one change has a predeclared comparable checkpoint and reversal condition.
If any check fails, the right Coach answer is not a stronger story. It is a smaller claim—or Not enough verified evidence.
