Cold Start Handling For Newly Listed Instruments

When to Use

Invoke this skill when an instrument must be sized and risk-managed before it has enough history to estimate from: a recent IPO or direct listing, a spin-off, a newly listed token, a new futures contract month, or a symbol that has just re-listed. A 30-day rolling volatility over 5 observations is not a small-sample estimate of volatility — for i.i.d. normal returns the sample variance from nu = n - 1 degrees of freedom has relative variance 2 / nu, so five days give a variance estimate with roughly 70% relative standard error. Fed to a volatility-scaled or Kelly-style sizer, that noise becomes leverage.

ColdStartHandler answers two questions that are deliberately kept apart:

1. What volatility should I use? The sample variance is blended toward a peer-group prior with the conjugate weight w = nu / (nu + nu_0), applied to variances. nu_0 ("prior strength") is the prior's weight in units of this instrument's own days.
2. How much capital may it take? A separate linear ramp, min(1, n_obs / warmup), times your base allocation. This is a risk-appetite policy, not an estimator.

When NOT to Use

• As a pre-trade risk control. This is a strategy-side sizing input that trusts its own configuration. Where SEC Rule 15c3-5 or an equivalent applies, the hard limits must sit in the broker-dealer's automated pre-trade layer; put this in front of that, never instead of it. See sec-rule-15c3-5-risk-controls-us.
• When no defensible peer group exists. Shrinking toward an arbitrary prior is worse than declining to trade the name: the output looks confident and is not. A newly listed token with no comparable, or a first-of-its-kind structure, has no prior worth using.
• When the short history is not the binding constraint. If the instrument cannot be borrowed, has no locate (us-reg-sho-short-sale-locate-requirements), trades in a 10-lot a day, or is still inside its first-day auction, sizing is the wrong lever. See liquidity-adjusted-position-sizing.
• For the model's features, not its risk inputs. Cold-starting a learned signal for a new symbol is a different problem — see categorical-feature-encoding-for-instrument-identity for identity encoding and transfer-learning-across-correlated-instruments for borrowing a correlated instrument's fitted structure.
• When the sample is short for a reason other than newness. A halted, suspended, or thinly traded old instrument has a short usable sample but not a cold start; treat stale data as stale, not as young.

Prerequisites

• A count of usable return observations, not a calendar difference. Listing date minus today over-counts: halts, no-trade sessions and missing bars are not observations. See references/workflows.md.
• A peer prior that is a single-name volatility, not a sector-ETF volatility. An ETF is diversified; its volatility is systematically below that of its constituents, so using it as the prior shrinks a new listing toward a number no single stock realizes. Use the median (not mean) realized volatility of comparable single names.
• Sample volatility and prior in the same units — both annualized, or neither. The module cannot detect a units mismatch; it will happily blend a daily sigma with an annualized one.
• A nu_0 chosen from the prior's own uncertainty, not by feel: for a prior whose estimate of sigma**2 has relative variance v, nu_0 = 2 / v.
• A warmup window justified by something. Market-structure anchors beat round numbers: IPO lock-up expiry (typically 180 days, contractual), index seasoning (S&P U.S. indices: 12 months on an eligible exchange). See references/standards.md.

Workflow

1. Count usable observations. Count bars actually present with valid prices. Exclude sessions the instrument was halted for the whole of, and exclude the listing auction itself — the IPO cross is a price discovery event, not a return.
2. Select the peer group before you need it. Same sector, comparable size, comparable float, and — where relevant — same recency-of-listing cohort. Freeze the selection rule; picking the peer group after seeing the instrument's early prints is a decision made on the data you are trying to shrink.
3. Refuse rather than default. If no prior is available, do not substitute zero and do not substitute the index. process_instrument raises on a missing or non-positive prior deliberately: a zero prior asserts a riskless instrument, and any volatility-scaled sizer divides by it.
4. Estimate with process_instrument(symbol, n_obs, observed_volatility, peer_prior_volatility). Below two observations there is no sample variance at all; the observed value is ignored (including if it arrives as NaN) and the prior is returned unchanged, with used_observed_volatility=False recording that.
5. Size against estimated_volatility, cap against max_position_cap_pct. They are independent controls. Note that the cap is a ceiling, not a target: a volatility-scaled sizer will usually ask for less during probation anyway, and the cap only binds when it does not.
6. Do not treat graduation as certainty. is_probationary going False means the size cap has reached the base allocation. It does not mean the volatility estimate is unshrunk — with nu_0 = 10, an instrument with 250 observations still carries about 4% prior weight, and that is correct. If you want the prior gone, lower nu_0.
7. Re-evaluate on every rebalance, not once at onboarding. Both outputs move with n_obs; a cached day-one decision keeps a stale cap for the rest of the ramp.
8. Watch for the events that reset the argument. Lock-up expiry, index addition, and the first earnings report all change the float and the volatility regime after the instrument has technically graduated. A 30-day warmup that ends before the first lock-up tranche unlocks has measured a float that no longer exists.

Full procedure: see references/workflows.md. Standards and sources: see references/standards.md. Printable pre-flight checklist: see assets/checklist.md.

Common Pitfalls

• Blending standard deviations instead of variances. w * sigma_obs + (1 - w) * sigma_peer is not the conjugate posterior and, because the square root is concave, it understates volatility whenever the sample and prior disagree. For a 5-day IPO at sigma_obs = 0.80 against a 0.20 prior, at the same weight w = 4/14, the standard-deviation blend returns 0.37 where the variance blend returns 0.46; the pre-2.0 module returned 0.30, lower still because it also weighted the sample by n / warmup. A risk control should not be biased toward the low side. The standard-deviation blend remains reachable via shrink_in_variance_space=False, for comparison during migration only.
• 0.0 * float("nan") is nan, not 0.0. The natural "weight the sample by zero when there is no sample" implementation emits NaN into the sizer for exactly the instrument the skill exists for. Absence of a sample has to be a branch, not a weight.
• Using a sector ETF as the single-name prior. Diversification makes the ETF's volatility structurally lower; every new listing is then shrunk toward a number no constituent realizes, and probation systematically over-sizes.
• Letting shrinkage end abruptly at the warmup boundary. A weight of n / N says a 29-day estimate is 97% trustworthy and a 30-day estimate is exact. Neither is true, and the jump puts a discontinuity in the sizing of every instrument on its graduation day.
• Counting calendar days as observations. IPOs halt — LULD trading pauses in the first sessions are routine — and a halted session contributes no return. Calendar counting graduates the instrument early on data it does not have.
• Treating the first-day price range as volatility. The IPO cross and the first session's price discovery are not draws from the return distribution you are trying to estimate; including them inflates the sample and, once weighted, the blend.
• Dropping newly listed names entirely for 60 days. The opposite failure: a hard exclusion forgoes post-IPO drift and liquidity-driven opportunity and, worse, hides the onboarding path from testing until the day it silently switches on. Use probation_floor_pct to throttle rather than exclude.
• Forgetting the borrow side. A newly listed name is frequently hard or impossible to borrow; a shrunk volatility says nothing about whether the short leg is executable.

Verification

• Run python -m unittest discover -s skills/cold-start-handling-for-newly-listed-instruments/scripts (29 tests), or python tools/run_all_tests.py for the whole repo.
• Documented example: ColdStartHandler(warmup_period_days=30, prior_strength_days=10.0) with n_obs=5, observed_volatility=0.80, peer_prior_volatility=0.20 returns confidence_weight = 4/14 = 0.2857, estimated_volatility = sqrt((10*0.04 + 4*0.64) / 14) = 0.4598, and max_position_cap_pct = 5/30 = 0.1667.
• Feed n_obs=0 with observed_volatility=float("nan") and confirm the returned volatility is exactly the peer prior and used_observed_volatility is False.
• Sweep n_obs from 0 upward and confirm max_position_cap_pct is non-decreasing, never exceeds the base allocation, and reaches it exactly at warmup_period_days.
• Confirm a zero or NaN peer prior raises rather than returning a zero or NaN volatility.

Related Skills

• categorical-feature-encoding-for-instrument-identity
• transfer-learning-across-correlated-instruments
• dynamic-position-sizing-based-on-realized-volatility
• liquidity-adjusted-position-sizing
• instrument-universe-change-detection-and-alerting
• point-in-time-index-constituent-tracking
• us-reg-sho-short-sale-locate-requirements
• new-strategy-onboarding-checklist