Auction Imbalance Feed Latency and Stale-Signal Slippage Playbook

Date: 2026-03-07
Category: research (execution / slippage modeling)

Why this playbook exists

Auction execution models often assume imbalance messages are effectively real-time. In production, they are not.

During open/close auctions, imbalance feeds can suffer:

• variable transport delay,
• bursty delivery (many updates arriving together),
• temporary sequence gaps or parser lag,
• clock misalignment between exchange timestamps and local decision clocks.

When execution logic reacts to stale imbalance snapshots as if they were fresh, desks pay a hidden cost I call Stale-Signal Chase Tax (SSCT).

Core failure mode

For a buy-side close participation strategy:

1. Router receives an imbalance update indicating buy pressure.
2. Message is already old by the time decision is made.
3. Strategy increases urgency (join higher / cross early / reduce passive wait).
4. True imbalance has partially mean-reverted (or flipped).
5. Desk overpays into transient pressure, then sees post-uncross reversal.

Symmetric for sells.

Key point: this is not pure forecast error. It is often latency-conditioned action error.

Data contract (minimum)

Per imbalance event and per child-order decision:

• symbol, venue, auction_session (open/close/reopen)
• imb_msg_seq, imb_event_ts_exchange, imb_recv_ts_local
• imb_value, imb_side, indicative_px, paired_qty, ref_px
• decision_ts, decision_action, decision_size, urgency_state
• order_send_ts, ack_ts, fill_ts, fill_px, fill_qty
• local book features at decision: spread, top depth, microprice, cancellation intensity
• clocks/offset metadata: clock_offset_estimate, source NTP/PTP state

Without imb_event_ts_exchange and decision_ts, staleness cannot be separated from market randomness.

Metrics that expose stale-signal slippage

1) Imbalance Signal Age (ISA)

[ ISA = t_{decision} - t_{imb,event} ]

Track p50/p90/p95 by symbol × venue × auction phase.

2) Freshness-Weighted Confidence (FWC)

[ FWC = \exp(-ISA / \tau) ]

τ is the regime-specific half-life of imbalance predictive value.

3) Stale-Chase Rate (SCR)

[ SCR_{\theta} = \frac{#{\text{aggressive actions with } ISA > \theta}}{#{\text{all aggressive auction actions}}} ]

High SCR means urgency is being driven by old information.

4) Feed Jitter Burst Index (FJBI)

Let (\Delta r_i) be inter-arrival gaps of imbalance messages at local receiver:

[ FJBI = \frac{\text{p95}(\Delta r)}{\text{median}(\Delta r)} ]

Large FJBI indicates bursty feed behavior where latest message age can jump abruptly.

5) Stale-Signal Chase Tax (SSCT, bps)

[ SSCT = 10^4 \cdot \frac{\sum_i q_i \big(p_i^{actual} - p_i^{cf,freshness_gated}\big)}{\sum_i q_i p_i^{cf,freshness_gated}} ]

Counterfactual uses same constraints but with age-aware gating (no stale-triggered aggression).

Modeling blueprint

Model auction imbalance signal as delayed observation of a latent process.

Latent state

[ I_t = \rho I_{t-1} + \epsilon_t ]

where (I_t) is latent true imbalance pressure.

Observation with delay

[ \tilde{I}{t} = I{t-d_t} + \eta_t ]

d_t is random delay (network + parser + scheduling), not constant.

Action-value branch model

Expected cost for each action (a \in {passive,join,cross}):

[ \mathbb{E}[C \mid x_t, ISA_t, a] = \alpha_a + f_a(x_t) + g_a(ISA_t) + h_a(ISA_t \times stress_t) ]

where x_t includes book pressure, spread, queue estimates, and time-to-uncross.

Policy objective

[ a_t^* = \arg\min_a \Big( \mathbb{E}[C\mid\cdot] + \lambda \cdot \text{CVaR}_{95}(C\mid\cdot,a) \Big) ]

with explicit staleness term in both mean and tail components.

Control design

Control 1) Age-gated urgency

Do not allow stale imbalance alone to trigger urgency step-ups.

• ISA <= θ1: full signal weight
• θ1 < ISA <= θ2: down-weighted signal
• ISA > θ2: imbalance cannot independently trigger aggressive mode

Control 2) Freshness-decayed blending

Use:

[ I_t^{used} = FWC \cdot \tilde{I}_t + (1-FWC) \cdot I_t^{book-proxy} ]

where book-proxy comes from current on-book pressure metrics.

Control 3) Burst-safe debounce near uncross

If time_to_uncross is small and FJBI spikes, require signal persistence for k updates before aggression escalation.

Control 4) Clock-drift guardrail

If clock integrity degrades (offset uncertainty beyond threshold), automatically widen stale thresholds or move to SAFE mode.

Control 5) Venue/session-specific freshness map

Calibrate θ1/θ2/τ separately for:

• open vs close,
• high-cap vs mid/small-cap,
• rebalance days vs normal days,
• each venue/feed path.

Calibration workflow

1. Reconstruct event timeline with exchange-event and local-receive clocks.
2. Estimate delay distribution d_t by phase/regime.
3. Fit action branch-cost model with ISA and interaction terms.
4. Build replay simulator with freshness-gated policy.
5. Shadow-run for 2-4 weeks, compare SSCT and completion risk.
6. Canary rollout with automatic rollback thresholds.

Promotion gates (example)

Promote only if canary shows:

• SSCT reduction >= 12%
• p95 auction slippage improvement >= 3 bps
• SCR reduction >= 15%
• no completion ratio degradation > 1.0pp
• no reject-rate increase > 0.3pp

Rollback if any two of:

• p95 slippage worsens > 5 bps for 2 consecutive sessions
• underfill breach exceeds tolerance
• SAFE-mode triggers exceed expected envelope

Common false conclusions

1. "Imbalance feed is exchange-native, so it is fresh enough."
   Exchange-native does not mean zero end-to-end delay.

2. "More frequent updates eliminate staleness risk."
   Bursty delivery can increase stale-trigger mistakes if policy ignores age.

3. "Auction reversal after fill is just market noise."
   Often partially driven by stale-signal chasing.

4. "One global staleness threshold is fine."
   Freshness half-life differs by session, symbol class, and event regime.

Minimal policy sketch

on imbalance_update(msg):
  update latest_imbalance and recv timing

on decision_tick():
  ISA = now - latest_imb_event_ts
  FWC = exp(-ISA / tau_regime)

  if ISA > theta2:
    stale_gate = true
  else:
    stale_gate = false

  I_used = FWC * imbalance_signal + (1-FWC) * book_proxy_signal

  if stale_gate and urgency_increase_requested:
    require persistence_k or deny escalation

  choose action via mean + tail objective

Desk-level takeaway

In auction execution, imbalance is not a timeless truth; it is a time-decaying signal.
If policy treats stale updates as fresh, slippage tails widen in exactly the sessions where desks care most (open/close/event auctions).
Model signal age explicitly, gate urgency by freshness, and measure SSCT as a first-class execution KPI.