Weak interpretation
“The strategy wins often, so I approve it.”
This skips payoff sizes, costs, validity, uncertainty and the decision constraints.
Level III - Manually evaluated assessment
Final Applied Quantitative Research Project
Interpret quantitative research evidence, identify methodological weaknesses, evaluate risks, and produce a defensible research recommendation.
Prepare your report using the research dossier and published assessment criteria. Revision and appeal windows begin after feedback on a submitted report; a local file check does not begin the grading period.
Northbridge research committee: reversal, trend and deployable evidence
Official assignment dossier · version 2026-10-17-v1
Northbridge is a fictional investment organization. Every strategy, name, price, cost and result below is a constructed assessment exhibit, not a real investment recommendation or performance claim. No external market data or programming is needed.
You are reviewing a proposed 10% allocation of a hypothetical 100 million portfolio. The existing portfolio would retain 90 million; the candidate sleeve would receive 10 million of investor capital. Candidate returns and all cost figures below are already measured on allocated investor capital. Each candidate targets gross notional equal to twice its allocated capital. Do not apply that leverage or turnover a second time to the supplied returns or costs.
The committee requires a credible expected incremental net benefit of at least 10 basis points per month for a normal production allocation, acceptable portfolio risk, feasible financing and exits, and an auditable validation process. A smaller, explicitly budgeted research or shadow pilot may be recommended to resolve uncertainty; it is not production approval or evidence that the minimum benefit has been met.
Your recommendation may be approve, conditional research/pilot, revise or reject. Explain the conditions and opportunity costs of your choice. You are not expected to rebuild a backtest from unavailable observations or invent probabilities for missing evidence.
Use a calculator or spreadsheet if helpful. No programming, external dataset, investment account or software purchase is required.
Research materials
These seven exhibits are the complete supplied evidence. Treat missing information as unknown; state any additional assumptions clearly.
| Item | Candidate R: residual reversal | Candidate T: trend continuation |
|---|---|---|
| Idea | Buy recent residual losers; short recent residual winners | Continue medium-term relative trends with a short recent-return skip |
| Feature formation | Final close, same-day market/sector residual and volatility estimate | Inputs observable by the previous close; rolling historical exposure estimates |
| Order assumption | Same closing auction; order cutoff 15:50 | Next-session participation schedule after signal formation |
| Data log | Final close available 16:08; residual file complete 16:12 | Release-aware files stored before orders; versioned snapshots retained |
| Exposure estimation | Market/sector coefficients fitted using the entire 2018–2025 sample | Coefficients refitted on eligible trailing observations only |
| Universe | June 2026 surviving membership applied back to 2018; delisted firms omitted | Historical membership and documented delisting treatment |
| Daily turnover | 90% of gross notional, one-way convention | 35% of gross notional, one-way convention |
| Record | Supplied information |
|---|---|
| R research search | 144 combinations of residual window, holding period and liquidity screen; only the best chart was included in the first pitch |
| R inference claim | Reported p=0.003 uses an independent-observation standard error on overlapping three-day holding-period outcomes; no selection correction supplied |
| R pitch | The author calls the historical result reliable causal mean reversion and requests the full allocation |
| T development | 2018–2024 development; eighteen specifications compared using 2025 outcomes |
| T 2025 claim | The author labels the selected 2025 result an untouched final holdout |
| T frozen version | Final chosen rule and preprocessing frozen on 31 December 2025; no 2026 retuning in the supplied log |
| T prospective evidence | January–June 2026 shadow forecasts, not live executed trades; simulated costs at a 5 million allocation |
| T forecast appendix | A new classifier improves validation Brier loss from 0.248 to 0.244 after comparing twelve feature sets on the same 2025 period; no separate prospective classifier comparison is supplied |
| Monthly figure, basis points of capital | R | T |
|---|---|---|
| Gross mean benefit | 72 | 36 |
| Estimated execution cost | 16 | 18 |
| Estimated borrow cost | 6 | 6 |
| Estimated financing cost | 4 | 4 |
| Pitch headline labeled net | 56 | 8 |
| Other exclusions | Taxes and management fees excluded | Taxes and management fees excluded |
| 2026 month | Gross benefit, bp | Execution, bp | Borrow, bp | Financing, bp |
|---|---|---|---|---|
| January | 40 | 12 | 5 | 4 |
| February | −20 | 12 | 5 | 4 |
| March | 60 | 12 | 5 | 4 |
| April | 10 | 12 | 5 | 4 |
| May | −30 | 12 | 5 | 4 |
| June | 48 | 12 | 5 | 4 |
| Item | Supplied information |
|---|---|
| Six-month calculation convention | For an illustrative standard error only, treat the six monthly net observations as independent; use sample standard deviation and s/sqrt(6) |
| Reference interval | A simple normal-multiplier interval for the six-month mean net benefit is approximately [−32.85, 26.85] bp using 1.96; this is a teaching approximation, not a verified small-sample confidence guarantee |
| Dependence limitation | The desk has not established independence, normality, stable costs or a stable future distribution |
| Existing portfolio ordinary volatility | 400 bp per month; T sleeve 300 bp; ordinary correlation −0.15 |
| Stress dependence scenario | Both volatilities become 600 bp per month and correlation becomes +0.70; no probability is supplied for this scenario |
| Historical drawdown evidence | T's selected historical sample has 7.8% observed maximum drawdown; only six prospective months are supplied |
| Risk mandate | A planned allocation must not rely on ordinary correlation as a guarantee of stress diversification. No supplied statistic proves a permanent 10% drawdown bound |
| Scenario | Capital | Gross monthly bp | Execution bp | Borrow bp | Financing bp |
|---|---|---|---|---|---|
| Small research allocation | 1 million | 36 | 8 | 6 | 4 |
| Intermediate allocation | 5 million | 36 | 12 | 6 | 4 |
| Requested allocation | 10 million | 36 | 18 | 6 | 4 |
| Large allocation | 20 million | 36 | 29 | 6 | 4 |
| Adverse case at requested allocation | 10 million | 26 | 30 | 12 | 8 |
| Item | Supplied information |
|---|---|
| Borrow audit | Ordinary inventory estimates cover 80% of proposed short notional. No committed inventory or recall stress is supplied for the other 20% |
| Liquidity | Several positions share crowded industry exposures. Calm-period volume and cost assumptions have not been validated during forced exits |
| Shadow execution | No real orders, realized slippage or fill rejection history exist for the prospective period |
| Operating controls | Model version and input logs exist for T; no approved feature-contract failure fallback, capital scaling gate or rollback authority has been documented |
| Unknowns | No scenario probabilities, future regime frequencies, corrected R result or full-scale live cost record are supplied |
Questions your report must address
- Define the actual decision and benchmark. Separate validity, statistical evidence, net economics, portfolio contribution and operating readiness. Prioritize issues by whether they can invalidate or change the action.
- Audit R's timeline, exposure estimation and universe. Show where unavailable information or selection enters. Specify the data and reconstruction needed; do not invent a corrected return.
- Reconcile both net headlines in Exhibit C. Calculate T's net benefit at every scale and in the adverse case. Convert the requested-allocation base and adverse monthly net benefit to cash. State the unit, capital denominator and exclusions.
- Calculate the six monthly shadow net observations and their mean. Interpret Exhibit E's approximate interval, its assumptions and its limitations. Explain why six shadow months are not equivalent to a long, executed live record.
- Evaluate the research search and the 2025 holdout claims. Address overlapping outcomes, selection and the classifier appendix. Propose a frozen, information-aware next evaluation without silently reusing selection evidence as a fresh test.
- At 90% existing portfolio and 10% T, calculate ordinary portfolio volatility using Exhibit E. Repeat under the stress covariance scenario. Explain why these conditional calculations and the historical drawdown do not guarantee future risk limits.
- Assess scale, crowded exits, incomplete borrow evidence and missing operating controls. Specify the implementation checks and safeguards needed before a pilot or production decision.
- Give one defensible recommendation with evidence-linked conditions. List the most decision-relevant unknowns, what would change your view, exposure limits and a monitoring/escalation plan. Distinguish supplied facts, calculations and your own assumptions.
Expected deliverables
- One PDF report, maximum 3 MB, with a concise executive recommendation and clearly labeled supporting calculations.
- A traceable exhibit-to-conclusion argument, including arithmetic, methodological limitations, implementation risks and conditional next steps.
- A source and assistance declaration. A calculator or spreadsheet is permitted; Python, external datasets and code are not required.
Aim for a concise research report of approximately 4–6 pages; this is guidance, not a required page limit.
Suggested report structure
- Decision and executive recommendation
- Evidence and statistical interpretation
- Validity and research-design audit
- Net economics, portfolio and implementation analysis
- Conditions, monitoring, unknowns and next evaluation
- Sources, assumptions and assistance declaration
Evaluation and passing standard
Quant Atlas reviews the report manually. Each category is rated from 0 to 4; its weighted points are category weight × rating / 4. A reviewer must explain each rating using your report and the supplied evidence.
| Criterion | Weight |
|---|---|
| Statistical interpretation | 25% |
| Identification of methodological flaws | 25% |
| Risk and economic evaluation | 20% |
| Quality of research recommendation | 20% |
| Clarity and justification | 10% |
Passing requires at least 75/100. At least 2/4 in both statistical interpretation and identification of methodological flaws, as well as at least 75/100 overall.
- 0 · Unsupported or materially incorrect
- 1 · Partial; major gaps remain
- 2 · Basic defensible treatment with limitations
- 3 · Sound, coherent and well supported
- 4 · Exceptional precision, prioritization and justification
A defensible conditional pilot, revision or rejection can earn full credit. Approval is not the expected answer by default. Unsupported certainty, invented evidence and ignoring material timing or cost problems weaken the assessment. Rounding differences are judged by their materiality and whether the units and reasoning remain clear.
Examples of reasoning quality
These separate miniature examples illustrate quality; they are not model answers to the Northbridge dossier.
Stronger interpretation
“In this separate illustration, 40 bp gross minus 45 bp execution cost leaves −5 bp before any omitted expenses. A high hit rate does not reverse those economics. I would investigate whether a feasible, frozen revision changes net value before approving exposure.”
This connects an explicit calculation to a limited recommendation and a decision-relevant next step.
Originality, sources and AI-use rules
Write the research reasoning and recommendation yourself. Disclose sources and assistance. Calculators, spreadsheets, accessibility tools, and spelling/grammar support are permitted. Do not submit generated analysis, fabricated exhibits, another person's report, or an undisclosed AI-written recommendation. An integrity concern requires human review and an opportunity to respond; no automated detector establishes misconduct.
At the end of your report, list external sources, tools used and their role. State “No external assistance” when appropriate. Accessibility assistance is not a reason for a lower grade. No particular citation style or paid tool is required.
Submission format
Prepare one PDF, no larger than 3 MB. Check the file requirements below, then contact Academy for report delivery instructions. The file check itself does not submit your report.
Report file check
Prepare your research report
PDF only · maximum 3 MB. Selection checks the file type, size, and PDF header in your browser. No file is transmitted or persisted.
This checks your file locally. Follow the submission instructions from Academy to deliver your report for grading.
Get submission instructionsGrading procedure
Projects are manually reviewed by Quant Atlas. Please allow up to 14 calendar days from successful submission for grading.
The review period starts only when live submission is open and a report has actually been accepted. Feedback will include criterion ratings, the weighted total, the decision and specific reasons. A selected local file has not been submitted or graded.
Revision and assessment review
Revision allowance
One standard revision is permitted after a not-passed result, responding to the reviewer feedback. The usual revision window is 30 calendar days after feedback; request an extension for accessibility or a documented technical problem. A further revision requires Quant Atlas guidance and explicit authorization.
Appeal process
Request an assessment review within 14 calendar days after feedback, identifying a specific grading or procedural issue and the relevant report/exhibit. Quant Atlas records the request and assigns a reviewer who did not make the original decision where available; any inability to provide a different reviewer must be disclosed. The outcome includes a written rationale. An appeal is not a new report submission and does not require a promise of a higher grade.
Use Academy support for guidance and review requests. Include your assessment reference and the issue you would like reviewed.
Certificate eligibility
The Certificate in Applied Quantitative Research requires passing the official Level I and II examinations, passing the Level III official examination, and passing this manually reviewed project under the applicable program requirements. Purchasing a level or preparing a PDF does not earn a credential. Successful graduates receive their certificate and the complimentary Ultimate benefit under the program terms.