Live Research Workshop

Is:

• A structured pitch of your research idea (15 minutes)
• An opportunity to get feedback before submitting the full IP
• A validation that your hypothesis is clear and testable
• A checkpoint to catch infeasible research early

Is NOT:

• A formal presentation (no slides required, no dress code)
• A paper or written submission (verbal is fine)
• A complete backtest (preliminary signal test is enough)
• A final judgment (feedback is iterative)

Allocate time as follows:

• Hypothesis (3 min): What is the edge? State the economic mechanism. What does the signal predict? In what direction?
• Data (3 min): What data do you need? Is it available? Have you confirmed with QD? Any lookback history concerns?
• Signal preview (5 min): What does the preliminary signal test show? IC? Directional confirmation? Regime dependence? This is NOT a full backtest — proof of concept.
• Next steps (2 min): What would the full IP contain? What's the biggest risk to the hypothesis?
• Q&A (2 min): Senior QR members ask hard questions. Be ready to defend.

• "Why hasn't this been arbitraged away?" If the edge is obvious, where are the competitors? If no one else is trading this, why not? What barrier to entry or frictional cost protects your edge?
• "What is your null hypothesis, and what result would falsify it?" You must have a specific quantitative threshold, not "seems promising."
• "How sensitive is the result to the lookback window?" Did you optimize this parameter? Test it over a range (20-day, 30-day, 60-day). If results are robust, say so. If they're fragile, acknowledge it.
• "What is the transaction cost estimate, and does the edge survive it?" Estimate bps round-trip. Subtract from your Sharpe. If Sharpe drops > 20%, edge is marginal.
• "What model assumptions are you relying on, and have you tested them?" You should have preliminary results from Week 3 assumptions testing already.
• "How does this strategy perform in 2008? In 2020? In rising rate environments?" Test on major market dislocations. If your edge breaks down in crisis, say so and explain why.
• "Is there any look-ahead bias in your preliminary test?" Walk us through your data pipeline. How did you avoid using future information?
• "What is the Information Coefficient, and how does it vary over time?" If IC is noisy (high variance), the signal may not be reproducible.

IP READINESS CHECKLIST

Section 1: Hypothesis

• Economic mechanism is written in one sentence
• Predicted direction is specified before any data analysis
• Falsifiability condition is stated (specific quantitative threshold)
• The hypothesis distinguishes behavioral from structural edge (or explains why it's neither)

Section 2: Data

• Data source, frequency, and date range are fully specified in 2a
• Cleaning and feature construction steps are written out with exact formulas
• Lookback history confirmed (> 5 years, ideally > 10)
• QD approval (Section 2b) has been confirmed in writing

Section 3: Methodology

• Signal formula is written with all parameters defined (no vagueness)
• Entry, exit, and sizing rules are unambiguous
• All relevant model assumptions are listed and tested in 3b (normality, stationarity, autocorr, heteroscedasticity, multicollinearity)

Section 4: Results

• Sharpe ratio, annualized return, win rate, profit factor reported
• Equity curve chart and drawdown analysis included
• Walk-forward or hold-out period used for out-of-sample validation
• Transaction costs included in net returns (5bps minimum round-trip)

General

• Information Coefficient test completed (IC > 0.02 is promising)
• No look-ahead bias in signal construction (data available at decision time)
• Hypothesis specified BEFORE data analysis (not result-first)
• All major market dislocations tested (2008, 2020, rising rates)

Path 1: You're cleared. Senior QR members confirm the hypothesis is testable, the data is feasible, and the methodology is sound. You proceed to submit the full IP.

Path 2: You need to revisit something. Feedback on a specific area: hypothesis needs clarification, data section has a gap, assumptions haven't been tested thoroughly enough. You address feedback and resubmit for workshop round 2 (if needed) or proceed directly to IP.

Path 3: It's not ready yet. Fundamental issues: unclear mechanism, data infeasible, IC too low. You and a mentor identify what needs to change. This is not failure — it's early course correction. Some analysts need 2–3 weeks post-workshop to address feedback.

SAMPLE PITCH: NasaPowerCouncil Corn Strategy (15 minutes)

[0:00–3:00] HYPOTHESIS

"My hypothesis is that satellite-measured soil moisture anomalies during the corn growing season predict corn yield surprises before USDA reports them. Here's the mechanism:

Corn yields depend on growing season weather — specifically soil moisture and temperature. USDA releases yield forecasts monthly, and these are based on surveys of farmers and historical yield models. There's a 3–4 week lag between end-of-month conditions and the USDA forecast release.

NASA satellite data — specifically soil moisture from POWER API — is available daily and is free. It directly measures the same growing season conditions that determine yields. If satellite data shows soil stress (moisture deviation > 1 std from seasonal) during key growth stages, the actual yield will be lower than the USDA forecast was when they released it.

The market prices corn based on USDA expectations. When yields come in worse than expected, prices fall. My edge is: I see the warning signal in satellite data before the USDA report, and I can trade ahead of the market repricing."

[3:00–6:00] DATA

"Data: NASA POWER API daily soil moisture and temperature for Corn Belt counties (IL, IA, MN, MO) from 2000 to 2024. That's 24 years of data — two complete bull-bear cycles. The data is free and publicly available.

I've discussed with QD (Jane Smith signed off), and NASA POWER can be ingested via Python requests into TimescaleDB. No licensing cost, no infrastructure barriers.

I clean the data by: (1) removing days with cloud cover > 50% (sensor bias), (2) interpolating single-day gaps (weather station noise), (3) aggregating to county level using county centroid lat/lon. Features: GDD deviation from 20-year seasonal average, computed as 30-day rolling sum of (T_max + T_min)/2 - 10°C base, minus the 10-year average for that calendar window."

[6:00–11:00] SIGNAL PREVIEW

"Preliminary signal test (2015–2022 in-sample): Information Coefficient between satellite GDD deviation and corn futures returns over the next 5 trading days is 0.068. This is meaningful — it's in the range where we see real edges.

Direction confirmed: negative GDD deviation correlates with negative corn returns (lower yields = lower prices). Win rate on directional accuracy in the hold-out 2023 period: 57%. Profit factor: 1.24 (gross).

The biggest regime dependence: the edge is strongest in July–August (peak growth stages). April–June, the correlation is weaker (early season is noisy). September onwards, the edge is lost (yields are largely set by then).

Signal formula: z_t = (GDD_deviation_t - mean(GDD_dev_t-60:t)) / std(GDD_dev_t-60:t). Entry: z < -1.5. Exit: z crosses zero or 10 trading days, whichever first."

[11:00–13:00] NEXT STEPS

"Full IP will include: (1) assumptions testing — Jarque-Bera, ADF, Ljung-Box on residuals; (2) full backtest on 2000–2024 with walk-forward validation; (3) transaction cost sensitivity (assuming 5bps round-trip).

Biggest risk: the edge could degrade if the market learns to incorporate satellite data. But this is unlikely in the next 2–3 years — satellite data is not in standard market feeds, and processing it requires domain expertise."

[13:00–15:00] Q&A

[Senior QR: "What's the data latency? Can you trade on the same day?"]

[You: "NASA POWER data is available at 8am ET the morning after observation. So Tuesday's soil moisture is available Wednesday morning. The corn market closes at 2pm CT. This gives us morning-to-close trading window. Full day trade is not possible, but morning window is usable."]