NOMURA QUANT CHALLENGE 5.
0
Question 3
TASK 4 WRITE UP SUBMISSION
1. PnL Validation Curves
The six plots below show PnL_validation(θ) swept across the full range θ ∈ [0, 1] for every client and each
hedging window τ. The red dashed line marks the optimal threshold θ* that maximises validation PnL for that
client/τ combination.
Figure 1: τ = 5s — PnL vs θ for Clients A–F
Nomura Quant Challenge 5.0 — Task 4 Write-up | Page 1
Figure 2: τ = 10s — PnL vs θ for Clients A–F
Figure 3: τ = 15s — PnL vs θ for Clients A–F
Figure 4: τ = 20s — PnL vs θ for Clients A–F
Nomura Quant Challenge 5.0 — Task 4 Write-up | Page 2
Figure 5: τ = 25s — PnL vs θ for Clients A–F
Figure 6: τ = 30s — PnL vs θ for Clients A–F
Nomura Quant Challenge 5.0 — Task 4 Write-up | Page 3
2. Over-Externalization vs. Under-Externalization Trade-offs
The θ parameter controls the broker/market boundary: at θ = 0 every trade is internalised (netted within the book);
at θ = 1 every trade is immediately externalised (hedged in the market). The curves reveal a sharp, asymmetric
trade-off.
Under-externalization (θ too low, θ < θ*)
When the threshold is set too low, the desk internalises too many orders. Inventory accumulates and the desk
carries unhedged directional risk. The validation PnL is near zero across all clients for θ < 0.35, indicating that the
spread income from internalisation is insufficient to compensate for the adverse price moves sustained before
eventual hedging. The penalty worsens with larger τ: at τ = 30s, Clients E and F show PnL = 0 for almost the
entire range θ < 0.45, meaning internalisation at wide windows generates no usable edge.
Over-externalization (θ too high, θ > θ*)
Once θ exceeds θ* the PnL plateau is reached for well-behaved clients (A, B, C), since nearly all flow is
externalised and residual inventory is trivially small. However, for clients D, E, and F — who exhibit noisier, more
adverse flow — over-externalization at larger τ causes PnL to turn sharply negative. This is visible from τ =
10s onwards: Client F at τ = 30s reaches −14,000 PnL for θ > 0.5. The mechanism is that externalising marginal,
low-signal trades incurs excessive market-impact costs and crosses the spread repeatedly, destroying value.
Key insight
The optimal θ* sits at the inflection point where the marginal cost of external hedging equals the marginal
benefit of inventory reduction. For high-quality clients (A, B, C) this is insensitive to τ (θ* ≈ 0.48–0.53). For weaker
clients (D, E, F) the optimal threshold degrades rapidly with τ, and eventually the correct strategy is full
internalisation (θ* = 0) because no externalization policy generates positive expected PnL.
3. Global vs. Client-Specific Thresholds
Case for client-specific thresholds
• Flow quality heterogeneity: Clients A and B consistently show θ* ≈ 0.49–0.53 with PnL in the tens of
thousands, while Clients E and F require θ* < 0.50 at short τ and θ* = 0 at long τ. A single global threshold
would be mis-calibrated for at least one group at any given τ.
• Adverse selection differs by client: The negative PnL region for E and F at high θ indicates these clients
carry informed or mean-reverting flow that punishes aggressive hedging. Treating them identically to A
would bleed PnL.
• Risk capacity: The magnitude of optimal PnL varies by a factor of ~10 across clients (A: ~18,000 vs. F:
~140 at τ = 30s). Client-specific thresholds allow capital allocation proportional to each client's actual edge.
Case for a global threshold
• Operational simplicity: A single θ reduces system complexity and the risk of overfitting per-client
parameters to the validation period.
• Robustness: Client flow characteristics shift over time; a global threshold around 0.50 is approximately
optimal for A, B, and C across all τ and never catastrophically wrong for any client at short τ.
Recommendation
Client-specific thresholds are justified given the clear divergence in optimal θ* across clients (range: 0.00 to
0.67) and the severe PnL consequences of misassignment for clients E and F at longer τ. The preferred
implementation uses a τ-adaptive, client-specific lookup table of θ* values, recalibrated on a rolling window.
4. Final Test-Set PnL
Nomura Quant Challenge 5.0 — Task 4 Write-up | Page 4
The table below reports the test-set PnL for each client and τ using the optimal threshold θ* identified on the
validation set.
Client τ=5s τ=10s τ=15s τ=20s τ=25s τ=30s
A 9,755 11,462 12,811 15,602 17,632 18,791
B 8,391 9,247 10,498 10,623 12,002 12,690
C 7,904 8,674 8,430 8,491 9,345 8,862
D 4,326 2,858 1,857 −524 −491 −28
E 3,155 1,034 0 72 71 0
F 1,810 −139 −161 0 2 −40
Total 35,341 33,136 33,435 34,264 38,561 40,276
All PnL values rounded to nearest integer. Rows marked 0 indicate θ* = 0 (full internalisation) was optimal — no
externalisation is performed and PnL is effectively zero on the test set for that client/τ.
Observations
• Clients A, B, C show monotonically improving PnL as τ increases, peaking at τ = 30s. These clients have
high-quality, internalizable flow where longer hedging windows allow more netting and greater spread
capture.
• Clients D, E, F deteriorate sharply beyond τ = 10–15s. The flow becomes increasingly adverse over longer
horizons — a signal of informed or directional client activity that cannot be profitably internalised at wide
windows.
• Aggregate PnL is broadly stable at ~33,000–40,000 across τ, with a modest upward trend driven by A, B,
C. The cross-client diversification masks the severe per-client deterioration in the weaker group.
Nomura Quant Challenge 5.0 — Task 4 Write-up | Page 5