Alpha-Decay-Aware Slippage Optimization Playbook

Date: 2026-02-24 (KST)

TL;DR

Most execution models optimize cost only (spread + impact + fees). Most alpha models optimize signal quality only. Live trading fails in the gap between them.

This playbook treats execution urgency as a function of:

1. Alpha decay speed (how fast edge dies)
2. Slippage convexity (how fast cost explodes with urgency)
3. Tail risk budget (how much p95/p99 pain you can tolerate)

Goal: stop paying either hidden tax:

• too slow → alpha evaporates before fill
• too fast → impact/markout dominates

1) Core Objective: Optimize Net Realized Edge, Not Raw Slippage

For remaining quantity (Q_r) at time (t):

[ \max_{\pi_t} ; \mathbb{E}[\text{AlphaCaptured}(\pi_t)] - \mathbb{E}[\text{ExecCost}(\pi_t)] - \lambda_{tail} \cdot \text{CVaR}_{95}(\text{Shortfall}(\pi_t)) ]

Where policy (\pi_t) controls participation, aggressiveness, venue mix, and child-order spacing.

Practical decomposition

[ \text{NetEdge}{bps}(t) = \underbrace{\alpha_0 e^{-\Delta t/\tau\alpha}}{\text{remaining alpha}} - \underbrace{C(\text{POV}, r_t, \text{liq}, \sigma, \text{tox})}{\text{expected cost}} - \underbrace{\text{TailPenalty}{95}}{\text{survivability}} ]

• (\tau_\alpha): signal half-life proxy (or e-fold decay)
• (r_t): execution regime (normal/stress/shock)
• (\text{tox}): toxicity signals (markout, OFI, VPIN-like metrics)

If you only minimize cost, you can “win” slippage and still lose PnL because alpha expired.

2) Data Contract (Minimum Viable)

Per parent order:

• decision timestamp, first-send timestamp, first-fill timestamp
• alpha snapshot at decision (score, expected bps, confidence)
• alpha snapshots every N seconds until completion/cancel
• child-order log (price, size, side, venue, type, queue estimate)
• fill log + post-fill markouts (5s/30s/60s)
• L1/L2 microstructure context at send/fill
• realized shortfall breakdown:
  • spread/fee
  • impact (instant + transient)
  • delay/opportunity

Without this contract, you cannot calibrate alpha-vs-cost tradeoff honestly.

3) Two Surfaces You Must Estimate

3.1 Alpha Decay Surface

Estimate by signal family and regime:

[ \alpha_{rem}(\Delta t \mid z) = f_{\alpha}(\Delta t, z) ]

Context (z): volatility bucket, trend/chop, session phase, event windows.

Use robust options:

• exponential: simple and stable
• piecewise spline: better fit for non-linear early decay
• quantile decay (p50/p80): safer under heavy tails

3.2 Slippage/Fill Surface

[ C = f_C(\text{POV}, \text{urgency}, \text{liq}, \sigma, \text{spread}, \text{tox}, \text{venue}) ]

Include:

• expected cost (mean)
• p95/p99 shortfall
• completion probability by deadline

Ignoring completion probability creates fake low-cost policies that simply underfill.

4) Urgency Metric That Actually Works

Define live urgency ratio:

[ U_t = \frac{\text{Marginal Alpha Loss per minute}}{\text{Marginal Cost Increase per minute}} ]

Interpretation:

• (U_t > 1): alpha decays faster than cost rises → speed up
• (U_t \approx 1): balanced tradeoff
• (U_t < 1): cost convexity dominates → slow down/passive

Fast approximation (desk-safe)

[ U_t \approx \frac{\alpha_{rem}(t) \cdot (1-e^{-\Delta/\tau_\alpha})}{C(\text{POV}{hi}) - C(\text{POV}{lo})} ]

Track this every 5–15 seconds for intraday strategies.

5) 3-State Controller (Production Friendly)

State A — Harvest (U low)

• objective: cheap capture
• lower POV, passive-first, queue-priority tactics
• strict toxicity veto to avoid adverse fills

State B — Balance (U mid)

• objective: mixed schedule
• moderate POV, selective crossing on favorable microstructure
• venue weights by expected net edge (not fee alone)

State C — Salvage (U high)

• objective: capture decaying edge before expiry
• higher POV, more aggressive crossing, wider venue set
• hard cap by tail budget + kill-switch thresholds

Use hysteresis to avoid A↔B↔C flapping.

6) Calibration Loop (Weekly)

1. Re-fit alpha decay by signal cluster and regime
2. Re-fit cost/fill surfaces with recent microstructure
3. Backtest policy frontier: net edge vs p95 shortfall vs underfill
4. Promote parameters only if:
   • net edge improves out-of-sample
   • p95 shortfall within budget
   • completion SLA preserved
5. Drift alarms:
   • realized alpha half-life deviates > X%
   • policy slips into frequent Salvage state
   • model residuals trend by session/venue

7) TCA Dashboard: Add These 5 Panels

1. Alpha-at-send vs Alpha-at-fill decay (bps lost to delay)
2. Net edge attribution: alpha captured - cost - tail penalty
3. Urgency-state occupancy (A/B/C time share)
4. Completion vs cost frontier drift (weekly)
5. False urgency rate: aggressive switches that did not improve net edge

If TCA only reports implementation shortfall, you are blind to edge evaporation.

8) Common Failure Modes

• Stale alpha priors: half-life learned in old regime
• Cost-only controller bias: “great slippage, bad PnL”
• No underfill penalty: optimizer chooses fake cheap paths
• Venue overfitting: wins disappear after routing drift
• Tail neglect: average improves while p95 explodes

9) 30-Minute Implementation Plan

Week 1

• add alpha snapshot logging at decision/send/fill
• publish baseline decay curves by top 3 signal families

Week 2

• compute live urgency ratio (U_t)
• shadow-run A/B/C state classifier (no trading action yet)

Week 3

• enable controller with conservative bounds in one symbol bucket
• compare against control on net edge + p95 shortfall + completion

Week 4+

• expand universe gradually
• auto-recalibrate weekly; freeze if residual drift exceeds threshold

10) Decision Rule (Desk Card)

Before increasing urgency, ask:

1. Is alpha truly decaying now (not just old prior)?
2. Is marginal speed worth marginal cost at current regime?
3. Will p95 shortfall stay inside budget if stressed one notch worse?

If any answer is “no,” do not pay urgency tax.

Closing Note

Execution is not a post-signal logistics problem. It is a joint inference problem over alpha decay, liquidity state, and tail risk.

The win condition is not “lowest slippage.” It is highest net realized edge with survivable tails.