Capstone Overview and Requirements

The brief

Build a working research environment in AmiBroker, use it to investigate one trading idea of your own, and write a report that answers thirteen specific questions about what you found.

The report is the deliverable. Everything else — the dashboard, the exploration, the scanner, the charts, the backtest — exists to make the report possible and to make its claims checkable.

!Important

A negative result is a pass. If your investigation concludes that the idea does not survive costs, or that the evidence is too weak to distinguish it from noise, and you can show the work that established that, you have completed this capstone successfully.

If your report concludes that you found something profitable and cannot answer the thirteen questions, you have not.

What you are building

The workflow the nine components implement

1. 1Component 1 — Market dashboardWhat kind of market is this? A state, written down, before you look at anything specific.
2. 2Component 2 — Stock explorationOne row per symbol: liquidity, trend, momentum, volatility, relative volume, relative strength, rank.
3. 3Component 3 — ScannerWhich symbols met a documented setup on the last completed bar.
4. 4Component 4 — Multi-timeframe chartOne candidate across three horizons, with a measured proof of no look-ahead.
5. 5Component 5 — Real-time or replay workspaceWatching a decision point arrive. Alert → human review → decision.
6. 6Component 6 — The strategyThe full specification, written before any test is run.
7. 7Components 7 and 8 — Backtest and robustnessPortfolio backtest, execution audit, sensitivity, out-of-sample, walk-forward.
8. 8Component 9 — The research reportThirteen questions, answered honestly, with evidence attached.

Deliverables checklist

Nine components, and for each one a file plus a written note. Tick these off as you go.

#	Component	You produce
1	Market dashboard	An .afl chart formula, plus a note recording the universe your breadth composites were built from
2	Stock exploration	An .afl exploration, plus one saved output table with the date it was run
3	Scanner	An .afl scan, plus a written specification of universe, liquidity, regime, setup and trigger
4	Multi-timeframe chart	An .afl chart formula, plus the look-ahead proof exploration’s output
5	Workspace	Either the real-time or the replay formula, plus a decision log with at least twenty entries
6	Strategy	An .afl file whose header states hypothesis, universe, entry, exit, stop, sizing, constraints and costs
7	Backtest	The backtest report, the trade list, three hand-checked trades, and the execution audit output
8	Robustness	The parameter surface, the out-of-sample result, the walk-forward result
9	Report	The thirteen answers, with evidence attached to each

Plus one thing that is not a component and matters more than several of them: a research log listing every specification you evaluated, including the ones you abandoned.

Assessment criteria

You are marking your own work, so the criteria have to be ones you cannot fool yourself about.

1. Could somebody else reproduce your result? Given your files, your universe definition, your date range and your settings, would they get the same number? If any of those four is missing, the answer is no.

2. Is every number attached to its assumptions? Universe, period, costs, fill assumption, data source, sizing model. A figure without those six is not evidence.

3. Did you validate the machinery before reading the result? Three hand-checked trades. The look-ahead proof. The execution audit. If you read the performance summary first, you have marked your own homework in the wrong order.

4. Did you write the specification before running the test? And is the file you shipped the one the specification describes?

5. Can you say how many specifications you tried? The number, from the log. Not “a few”.

6. Is the conclusion proportionate to the evidence? This is the hard one. A single backtest over one universe in one period supports a much smaller claim than most people make from it.

7. Does the report say what would change your mind? If no result would have, the study was decorative.

∴Research note

Criterion 6 is where nearly everyone loses marks, and the failure is not dishonesty. It is that after three weeks of work, “the evidence is inconclusive” feels like a failure rather than a finding.

The defence is mechanical: write down, before you run the backtest, what result would count as support and what would count as refutation. Then hold yourself to it. That sentence goes in the research log, dated.

The Level A path, in full

Every component works without a paid data subscription. Here is exactly what changes.

Component	Level A (free EOD data)	Level B/C (broker or vendor feed)
1 Dashboard	Unchanged. Daily bars, breadth composites built from your own EOD universe	Unchanged
2 Exploration	Unchanged	Unchanged
3 Scanner	Unchanged	Unchanged
4 MTF chart	Weekly context over daily bars	Optionally intraday context over minute bars
5 Workspace	Bar Replay drives the chart. GetPlaybackDateTime() labels the session	A live feed drives the chart; GetRTData() adds a quote panel
6 Strategy	Unchanged	Unchanged
7–8 Backtest	Unchanged	Unchanged
9 Report	Unchanged	Unchanged

✱AmiBroker Professional required

Component 5 is the only place the edition matters. Real-time data requires the Professional edition, and the GetRTData() reference states that calling it on Standard returns Null for every field.

The replay version is not a simulation of the real version — the decision logic is the same code. What Bar Replay removes is the pressure of a live market, and what it adds is the ability to step one bar at a time and write down why you decided what you decided. For learning, that is the better tool.

How long this should take

Component	Time on task
1 Market dashboard	60 min
2 Stock exploration	60 min
3 Scanner	55 min
4 Multi-timeframe chart	55 min
5 Workspace	55 min
6 Strategy	60 min
7–8 Backtest and robustness	70 min
9 Research report	60 min
Total	about 7 hours 45 minutes

Those figures assume a database already in place and the specification in Component 6 taken as written. Designing your own hypothesis, building a defensible universe and actually thinking about the results will take considerably longer — and that time is the point of the exercise, not overhead on it.

How to submit it to yourself

There is no marker. That is a problem, and the fix is to make the work reviewable by a version of you who has forgotten it.

Put everything in one folder, dated. Formulas, exported tables, backtest reports, the research log, the report.

Save your Analysis configurations as .APX files. The settings that live in dialogs — Apply To, Range, periodicity, report options — are exactly the half of a result you will fail to remember, and an .APX is self-contained.

Write the report first, in outline, before you have the answers. The thirteen questions are known in advance. Writing the questions down and leaving the answers blank makes it obvious which evidence you still need — and stops the report becoming a description of whatever you happened to find.

Then put it away for two weeks and read it again. Everything unclear on second reading was unclear the first time; you just knew what you meant. That second read is the closest thing to peer review you have.

▷Exercise

Before starting Component 1, write three things in your research log and date them:

1. The idea you are going to investigate, in one sentence, in plain language.
2. What result would make you abandon it.
3. The universe you will use, and how you built it — including whether it is a point-in-time list, and if not, why not.

Those three lines are the beginning of the report, and having them written down before you see any result is the single most effective defence against the failure modes in Part 30.

Recap

Nine components, one report, thirteen questions. The report is the deliverable and a negative result is a pass. Every component works on free end-of-day data, with Bar Replay standing in for a live feed in Component 5 — the only place the Professional edition matters. You mark your own work against seven criteria, of which the two that matter most are whether somebody else could reproduce your result and whether your conclusion is proportionate to your evidence. And the research log, started before Component 1 and dated, is what makes the sixth criterion answerable at all.

Check your understanding

Question 1. Your capstone investigation concludes that the idea does not survive realistic costs. Have you passed?

• No — a capstone must produce a working strategy
• Yes, provided you can show the work that established it and answer the thirteen questions
• Only if you then find a variation that does work
• Only if the backtest covers at least ten years

Show the answer and why

Answer: Yes, provided you can show the work that established it and answer the thirteen questions

The deliverable is a piece of research you would defend, not a profitable system. A well-evidenced negative result is a genuine finding; an unevidenced positive one is not.

Question 2. Which components change on the Level A (no subscription) path? Select all that apply.

• Component 5, where Bar Replay drives the chart instead of a live feed
• Component 4, where the context interval is weekly over daily rather than optionally intraday
• Component 7, because backtests need real-time data
• Component 9, because the report requires live results

Show the answer and why

Answer: Component 5, where Bar Replay drives the chart instead of a live feed, Component 4, where the context interval is weekly over daily rather than optionally intraday

Backtesting and reporting are historical activities and need no feed at all. Only the workspace has a genuinely different data source, and its decision logic is identical in both versions.

Question 3. Why does the overview insist on writing the report outline before you have the answers?

• To save time later
• Because the thirteen questions are known in advance, so an empty outline makes it obvious which evidence is still missing — and stops the report becoming a description of whatever you happened to find
• Because AmiBroker cannot export reports afterwards
• Because the report must be dated before the backtest

Show the answer and why

Answer: Because the thirteen questions are known in advance, so an empty outline makes it obvious which evidence is still missing — and stops the report becoming a description of whatever you happened to find

A report written after the results is shaped by them. A report structured before them is a list of things you have committed to being able to answer, which is a much stronger constraint on your own reasoning.

Question 4. Which of these are required before you may read the backtest performance summary? Select all that apply.

• Three hand-checked trades reconstructed against the chart
• The multi-timeframe look-ahead proof output
• The execution audit exploration output
• A profitable equity curve

Show the answer and why

Answer: Three hand-checked trades reconstructed against the chart, The multi-timeframe look-ahead proof output, The execution audit exploration output

All three validate the machinery that produced the number. Reading the summary first is marking your own homework in the wrong order — and a performance threshold is a result, never a validation step.

Question 5. What does the GetRTData() reference say about the Standard edition?

• It works but returns delayed values
• It is available ONLY in the Professional edition; calling it on Standard gives Null values for all fields
• It requires an additional plug-in licence
• It works only during market hours

Show the answer and why

Answer: It is available ONLY in the Professional edition; calling it on Standard gives Null values for all fields

That is the documented behaviour, and it is why Component 5 has two formulas. The replay version uses no real-time function at all, so it runs identically on either edition.

Check my answersClear