Education in research methods · Quantic Eagle operates only with its own capital. No client funds, investment advice or trading signals.

IN PREPARATION · FREE RESOURCES AVAILABLE NOW

Before Capital

Before trusting a backtest — a simulation of purchases and sales on past data — find out what its results actually establish.

Is a more complex model also a better choice? In the final project of “Before Capital”, you use invented prices to compare a model with a basic rule: always buying the same stock. The model loses €152.17, while the simple rule earns €35.73. You then increase execution costs, and the simple rule loses too. You learn to choose a benchmark, keep test conditions comparable and assess how much the result depends on costs.

Try the comparison below Explore the learning path

Profit rises from €294 to €448. Did the filter help?

In this invented example, the first simulation buys and sells on all 48 scheduled occasions, investing €1,000 per purchase. The second skips 12 occasions according to a fixed filter and invests €2,000 per purchase. Prices are unchanged; each sale takes place before the next purchase. Each completed purchase and sale costs €2 in total.

Net profit rises by about 52%, but you changed two things: which trades to take and how much to invest. Which comparison lets you assess the filter while keeping the amount invested the same?

Show the comparison at €1,000 per purchase

Simulation	Net profit
All 48 trades	€294
36 trades admitted by the filter	€188

With the filter and €1,000 per purchase, profit falls from €294 to €188. In this example, the filter reduces profit at the same investment amount. The €448 result also reflects the doubled amount; on its own, it does not show that the filter improves the strategy.

The data and filter are invented. This comparison isolates a difference in profit for this example; it does not measure every risk or establish future performance. QE-COMPARE-02 / 1.0.

Follow the calculations and compare all four simulations

Where are you starting from?

For readers building a quantitative research project. If you come from data science and are new to finance, the opening module covers the basics of working with market data.

You are designing your first system

You can work with data and want to organise your first trading research project, from understanding market data to comparing models and planning work for AI agents.

You already have a backtest or pipeline

You want to make better research decisions, compare models consistently and identify which parts of the process deserve more work.

You work independently or in a small team

You want to give AI agents precise assignments, make experiments reusable and measure the time, cost and quality of completed work.

What you learn to check in a backtest

You examine results that look reassuring and are not: two programs that agree on a commission a hundred times too small; closed trades with an overall profit while the account, counting positions still open, is down; a preliminary ranking that changes once orders and costs are simulated. Each time, you identify the check that would have caught it.

Check that the input was available

A simulation buys at 10 am using an input received at noon. Compare the decision time with the receipt time and prevent the model from using information it could not have known then.

Recalculate what a change affects

Correct a price used by the model: the old predictions no longer describe those inputs. Follow the calculations that depend on that price and regenerate them. If you change only a rule applied after prediction, check when you can retain the scores and rerun just the order simulation.

Explain whether to proceed, revise or stop

Separate results you have actually verified from open questions. Explain which observation supports your choice and what evidence is missing before committing capital. The dossier retains the data, versions and reasons for that decision.

Assess training stability

Training a network twice with the same data and settings can produce different models because the calculation starts from different random values. The exercise compares twenty training runs with those shared conditions and different seeds: seeds identify the random initialisations. Make the choice before seeing subsequent data and compare the selected model with a model group defined in advance. The best of twenty results on data already inspected is not confirmation on new data.

Recognise a winner created by selection

In an invented experiment with no advantage, choose the best of forty attempts over the same period. Then observe how that same winner performs on new data: on average, most of the apparent advantage disappears. Learn to distinguish the result used for selection from a subsequent evaluation.

The fieldbook assumes basic statistics and familiarity with tables of data. The learning path revisits the distinction between training, selection and evaluation data. The opening module introduces financial data; the final project offers a guided entry covering percentages, costs and reading the ledger. Coding exercises require basic Python and terminal skills; the free example needs only a browser.

The lessons we wish we’d had at the start.

The fieldbook draws on around ten years of accumulated research experience, with contributions from the founder and external collaborators. Successful experiments, mistakes and changes of direction become cases you can examine: what we tried, what we found and why we changed our approach.

If you understand percentages and tables but have gaps in your quantitative foundations, a guided entry and financial data module help you begin. Running the tools requires the Python basics stated in each activity. If you already run backtests, the final project gives you four versions of an experiment without saying which ones are broken. One is correct but loses money; the others hide an input assigned to a decision before it was available, predictions not updated after a change, or a selection period listed twice. You choose the checks, find each fault and explain why a correct losing result is still a valid result.

The experience described concerns research; it is not a record of trading performance.

Already have a backtest? Start with the lab cases and dossier to see which decisions change after a check and which questions remain open. Compare the expected choice with the program’s choice → See the decline the balance hides →

From an idea to a test

It worked in the past. What happens next?

Imagine an invented rule: buy after a daily price fall of at least 2% and sell after five trading sessions. Using prices from 2015 to 2017, you try different minimum falls, holding periods and sale rules until one combination makes money. You have selected settings with knowledge of that period’s results. You still need to evaluate them on prices that did not guide the selection.

2015–2017
Build the rule
Try thresholds and holding periods, then choose the settings that made money. Those prices have already shaped the strategy.
2018 · first test
Open the data you held back
The rule is already fixed. Observe what it does with prices you did not use to choose it.
2018 · after a change
The same year now has a different role
The result disappoints you, so you change a threshold and try again. You are now using 2018 to make choices; it is no longer an independent test.

The next test needs data that have not yet guided your choices

You can use 2019 for this only if you have not already inspected it to change or choose the strategy. Calling a file “test” or “out of sample” does not change how you used its data.

Why might it fail later? Among many combinations, some can profit from coincidences found only in the prices already inspected. Selecting those combinations can make a coincidence look like a useful rule. Markets and costs can also change. A disappointing result in a new period calls for investigation; it does not identify the cause on its own.

What to do in your own project

Write down the idea before seeing results
Write down why the rule might make sense and what would make you drop the idea. Record every version tried, including the losing ones.
Fix the comparison in advance
Use one period to build and choose the strategy. Fix rules, trade sizes and cost assumptions before opening the next period. The year 2019 can be a fresh test only if it has not already guided your choices.
Move forward through time
Repeat the process moving forward through time: at each date, use only what you could have known then. Keep the losing periods too, without selecting only favourable ones.
Find what the profit depends on
Check how much the result depends on one precise threshold, a few lucky trades or optimistic costs. A result that disappears under small, plausible changes needs investigation before you rely on it.

This is what we mean here by robustness: looking for results that depend on more than the fortunate details of one test. These checks help reject fragile ideas; they cannot guarantee that a strategy will make money in the future.

What do you test after the backtest?

The fieldbook includes seven activities on rankings, trading frequency, model groups, data checks, calculations to repeat, paired tests, and records. Each has invented inputs, a solution, and a worksheet. The three examples below preview that work.

01 · Choose a model
Profit on training data can distort the ranking
A earns a lot on the prices used to train it but loses on the next period. B earns less in the first period and more in the second. Recalculate the ranking to see which part of the formula rewards A.
Chapter 04 · Comparing candidates
02 · Assess risk
It has never lost. How often has it traded?
One strategy has three trades, all profitable; another has eighteen. Count days with open positions and compare observed losses before calling the first strategy safer.
Chapter 05 · Activity and losses
03 · Combine models
The strongest individual models can lose at the same time
A and B earn good results separately but lose at the same times. Try pairing A with C, which behaves differently. Calculate what happens when you split capital between them, then when you require their agreement before trading.
Chapter 07 and advanced module §9 · Working together

Explore the programme and exercises →

Before you trust the backtest.

Learn to spot the data mistakes that flatter a backtest, then investigate whether adding another model actually improves the system.

A starting point for data scientists new to finance.

The opening module explains how stock splits and dividends affect historical prices, when an input becomes available and which dates an average requires. It also shows which data transformations must be learned from the training period before being applied in the next evaluation.

How well do you know your model?

The advanced module separates two questions: what today’s model would decide on past data, and what decisions were actually possible at the time. Using invented data, you compare an unchanged model with a system that adds a second model, and examine both favourable and unfavourable outcomes.

These two modules are part of the fieldbook in preparation, with exercises using synthetic data, diagrams and worked answers. Quantic Eagle’s trading models and operating rules remain private.

After a test, you need to be able to explain to someone else how you obtained its result. The fieldbook “Before Capital” teaches you to design and compare trading experiments. Its Research Verification Protocol organises checks on data, decisions and results. The sample dossier distinguishes completed checks from those still needed before deciding how to proceed.

Twenty chapters and six workshops, plus four modules: financial data, evaluating a second model added to an already trained model, a model-selection lab and a final project. The package includes 26 worksheets, Python checks and synthetic examples you can reproduce.

Choosing a model: beyond the highest profit

Should you choose the model with the highest profit? In an invented exercise, the two periods have equal length and the same starting capital. X earns USD 500 in the first period and USD 30,000 in the second; Y earns USD 8,000 in each. Total profit favours X, but hides how much its result depends on the second period.

The lab’s criterion also considers balance between periods, which is why it prefers Y. In the exercise, both have 100 trades per period and a maximum drawdown of 12%. In the lab, you compare the criterion, explained in full in the private fieldbook, with total profit and the profit-to-drawdown ratio; you change one component at a time to understand how the ranking changes.

With the exercise settings, the criterion prefers Y to X. In the invented continuation, which was not used for selection, X earns USD 1,200 and Y loses USD 800. A more elaborate ranking does not guarantee the future winner: you learn to justify a selection, expose its assumptions and test the chosen candidate on subsequent observations.

A different case: Count improvements and positive profits in the free case

Six decisions for designing your research.

Want to turn a trading idea into a simulation you can check? The path follows six decisions: what to test, which data to use, what to compare it with, how to measure the account, how to assign analyses and how to justify the next step. Each decision has a worksheet, case or exercise.

A map of the design process. Arrows show the working sequence; an experiment may send you back to an earlier decision.

The idea

Write down the rule to simulate and why it might be useful. Choose a comparison: the previous version if you are changing it, or a simple rule on the same data. State what result would make you abandon the idea. Chapter 01 and worksheet 01.

The data

Before simulating orders, list the stocks eligible for purchase on each date and check when each input could be known. A price, financial statement or indicator updated today was not necessarily available then. Chapters 02–03 and worksheets 02–03.

The models

To check whether a change helps, run the old and new versions over the same period with the same data, capital and costs. If you add a model, also run the system without it. This avoids attributing a result caused by a different period or amount to the change. Chapters 04, 07 and 13.

The risk

Reconstruct account value throughout the simulation, including gains and losses on positions still open. Measure falls from earlier peaks and connect changes to orders and costs: final profit alone can hide losses you would have had to bear first. Chapters 05 and 11, account laboratory.

The AI agents

If you ask an AI assistant to examine research files, give each request a precise list. Check that its response covers those files and review the analyses, counting repeated requests and human review time. Here the AI agent uses a language model for analysis tasks, not trading operations. Chapters 08, 15 and 20 and workshop 26.

The next experiment

Collect the data used, commands run, results and limits of the checks in a dossier. It supports a decision to continue, correct or stop the research, and lets another person assess that decision. State the next test and what result would change your decision. Chapters 10 and 19, sample dossier.

Four checks before relying on a backtest.

The distinctive material is a selection of cases, decisions and practices drawn from internal research: useful experiments, abandoned approaches, corrections and findings that changed the design. The fieldbook explains the transferable reasoning and gives you ways to practise it.

Problem	What you can try now	What you learn to distinguish
Profit changes between two simulations	Guided case with three purchases	How different investment amounts change results while predictions stay the same
The program passes checks but gets a decision wrong	Three Museum cases	A check that passes from a result that is actually correct
You have an account-value series	CSV mapper and drawdown check	Losses visible in the data and unobserved movements
You need to choose the next step	Sample dossier	What you have verified, what you still need to test and what to rerun after a fix

An illustrated fieldbook you can read in a browser, with workshops, editable worksheets and tools to run on your computer. Start with the financial-data primer and return to the advanced material as your project develops.

Three assumptions worth testing.

Does the highest-ranked model also make the most money?

In the internal comparison of 192 candidates, a preliminary score helped decide which models to examine. The ranking changed when orders were simulated and economic results measured. The case teaches you to compare the two rankings and identify models the initial filter would have discarded. It does not establish which model will profit in the future.

Does adding a second model improve the first?

Compare a simulation of the base model with one that also uses the second model. Use the same period, capital and cost assumptions, and reconstruct the orders from both. You can then see whether the additional model changes purchases, results and losses during the test. The fieldbook also includes a case where it makes results worse.

Do identical predictions produce identical purchases?

Before replacing a function that decides orders, check that the rewrite follows the same rules. Two programs can both give A a score of 0.7 and B a score of 0.6, then buy different stocks or different quantities of the same stock. In the exercise, you compare their decision records and find the first stage that changes the decision, before looking at final profit.

Each case states its scope. Examples reconstructed for teaching are distinguished from internal observations; the lessons are not presented as universal laws of markets.

See the reasoning at work.

Open the Museum cases. In each one, click “Corrected case” to compare the original output with the result after the explained correction. You do not need to edit data or code to use the diagrams.

01 / EXHIBIT

Same scores. Different decisions.

You want faster backtests and rewrite the function that selects a stock to buy. The model scores each stock: A = 0.7 and B = 0.6. It has not selected a purchase; the next rule must buy only the stock with the highest score. In the rewrite, you reorder the stocks from A, B to B, A, but keep the scores in their original order. The program pairs them by list position: it now attaches 0.7 to B and 0.6 to A.

02 / EXHIBIT

Two approvals. Did the second reviewer get a fresh start?

An agent has prepared your backtest and you want a second opinion before trusting the profit. The second reviewer also finds the first agent’s favourable report in the folder. Are two approvals enough? For a fresh check, arrange a reading of the same originals without showing the other reviewer’s verdict first.

03 / EXHIBIT

The average uses the wrong dates.

You want to use a price average to decide a purchase. The program can return a plausible number even when the download is missing a session: check the dates used before passing the average to the model. At the opening of session 4, the available price is €10. The invented rule buys only if that price exceeds the mean closing price of the three previous sessions: 1, 2 and 3. Their original closes are €10, €11 and €12, giving a mean of €11 and no purchase. But session 2 is missing from the downloaded CSV. The function takes the last three prices present, using €5 from session 0 instead of the missing €11. The mean falls to €9.

The False Green Museum Open a sample dossier

A method you can put to work.

In the fieldbook in preparation: seven Python commands, each addressing a specific question:

equity:: measures account-value declines in the supplied observations.
temporal:: compares input availability times with decision times.
folds:: groups earlier data for training and later data for evaluation, excluding training rows whose outcomes are not known before evaluation starts.
ranks:: compares two rankings of the same models.
trace:: compares the declared decision records of two programs, from score to selected quantity.
batches:: compares work assigned to agents with results delivered.
windows:: looks for the observations required by a calculation’s calendar.

Practical workshops: 6 · Fieldbook chapters: 20
Open a sample dossier · Printable HTML
MT5 observer · teaching tool to test on a demo account for recording equity and balance
Agent templates in your chosen language: AGENTS.md / CLAUDE.md / .cursor/rules

The fieldbook teaches research design and verification. Developing your own system still requires data, implementation and experimentation; a broker-ready trading strategy is not included.

I only have a list of trades

A list of purchases and sales does not show what the account was worth while positions were still open. Start with the free case comparing balance and total account value. The planned fieldbook includes a workshop on starting to record these observations in a demo environment; the recorder cannot reconstruct the past.

I have an equity curve

Open the free mapper, confirm columns and cash flows, then use the checker. The measurement covers only available samples.

I am building a pipeline

Explore the Museum cases: compare two programs’ decisions, check which original files and earlier verdicts each reviewer receives, and locate a missing price. In the dossier you can also check duplicated file analyses. The planned fieldbook includes Python programs for repeating these checks on test files and adapting them to your data.

Open research desk

Read the research. Try the tools. Follow the fieldbook as it takes shape.

The article and browser tools below are free to use now. The complete fieldbook is in preparation; its final contents and any purchase terms will be published before orders open.

01
Inside the research lab
Follow a change from inputs to orders, distinguish computation from waiting time, and retain original decisions when rerunning the past.
02
Observed equity check
Measure drawdown from the observations in your CSV, with explicit limits.
03
Three interactive cases
Compare faulty and corrected cases to see what the first check did not verify.
04
CSV column mapper
Choose your columns and formats explicitly. Keep your data in your browser.
05
A worked decision dossier
See how findings, supporting files and missing checks fit together.

Two programs can agree and still be wrong

You want to choose a strategy to investigate further and look at the profit from a backtest, a simulation of purchases and sales using historical data. If the program deducts less commission than the assumed rate requires, a strategy can look better simply because its costs are wrong. Two programs can give the same result and still hide this problem if they share the function that converts the rate.

We start with a calculation you can do by hand. It helps you see what to check before trusting the calculations in a larger system. Work out the expected fee from the rate and the purchase amount, checking both separately from the program.

In this invented case, you buy shares worth €8,000 in total: this is the purchase amount, not a number of shares or lots. The assumed commission is 1 basis point, or 0.01% of that amount. The simulator records €0.008 and a second program returns the same number. Calculate the commission yourself before opening the answer.

Try the free case and check the calculations

The goal: a decision you can defend

In this project, a cost of 1 basis point is 0.01% of each order’s amount: 2 basis points mean 0.02%, and 8 mean 0.08%. “Per side” means the cost is applied separately to each purchase and each sale. The final project runs one complete invented experiment, from data to orders and costs. You fix the model before the final test and compare it with always buying the same stock and with holding cash. When execution costs rise from 2 to 8 basis points per side, even the simple alternative loses, and cash becomes the benchmark to beat. You also repair three broken versions of the experiment and run them again. The public sample dossier shows how to document a narrower case.

Examine the public dossier

Before requesting a release notification

Why sell a fieldbook if you work in trading?

Education is also a commercial activity for us. Some of the experience gained through research can become useful cases, exercises and tools while we keep our operational methods private. The fieldbook grows out of the parts of that work we can share.

Is this a course for beginners?

You need basic statistics and familiarity with tables to begin. The opening module and guided project help fill gaps in your knowledge of financial data and research methods. Coding exercises require the Python basics stated in each activity. Read the prerequisites →

Can I buy or reserve it now?

No. We are sharing free resources while preparing the fieldbook. Requesting a notification does not place an order, pay a deposit or reserve the fieldbook. Price and release date have not been announced.

Will it include Quantic Eagle’s trading strategies?

The material explains research methods, design choices and selected lessons from internal work. It does not include our proprietary buying and selling rules, trading signals, operational models or any promise of returns.

In which language will I receive the fieldbook?

Request the notification in English, Italian or Spanish. The planned delivery is one language per package; the final editions will be described before sales open.