Skip to content

Algo wheel

How an algo wheel allocates eligible orders among broker algorithms, measures the realised cost mix, and supports comparisons without proving a universal best algorithm.

An algo wheel is a rules-based mechanism that allocates eligible orders among a panel of brokers' execution algorithms. It makes routing repeatable and creates a record from which a trading desk can compare execution outcomes. An allocation rule can use rotation, weighted random assignment or performance-based weights; the word “wheel” does not imply that every order is sent randomly.
The practical quant problem is that a broker's measured cost depends on which orders it receives. A wheel can help control that assignment, but an automated allocation is not, by itself, evidence that the selected algorithm is best. The SEC staff's 2020 report, pp. 35–36 describes wheels as selecting brokers and algorithms under specified preferences or constraints and supporting performance comparisons across market conditions.

Eligibility comes before the spin

Consider a desk that wants to compare three brokers' VWAP algorithms for a defined segment of equity orders. The wheel first needs to establish which orders belong in that segment and which broker–algorithm combinations are permitted for each one. Region, instrument coverage, order size relative to available liquidity, price limits and urgency can change that eligible set.

This is a different decision from the execution algorithm's scheduling and venue choices. In the example here, the wheel assigns a parent order to a broker algorithm; that algorithm then manages its execution. Smart order routing may determine where individual child orders seek liquidity. A wheel need not replace those downstream decisions.

Suitability and objective also precede scoring. Yang and Jiu's Algorithm Selection: A Quantitative Approach, written by ITG researchers, separates whether an order is suitable for algorithmic execution from which algorithm to choose. Their framework conditions evaluation on order requirements, stock characteristics and market conditions, using information available at order entry. Comparing an urgent completion strategy with a patient VWAP strategy without acknowledging their different objectives can reward the wrong behaviour.

Ten assignments, one realised cost

Suppose all ten eligible parents have USD 1 million of intended notional, are fully completed, and use the same cost benchmark, horizon and fee convention. Positive cost means adverse execution. The following is a fictional realised sample, not broker-performance data:
Broker algorithmTarget assignment probabilityRealised parent countRealised notionalObserved average cost
A50%5USD 5m5 bp
B30%3USD 3m4 bp
C20%2USD 2m6 bp
A basis point is one ten-thousandth. A's currency cost is therefore USD 2,500, B's USD 1,200 and C's USD 1,200. Total cost is USD 4,900 on USD 10 million:
cˉ=5(5)+3(4)+2(6)10=4.9 bp.\bar c=\frac{5(5)+3(4)+2(6)}{10}=4.9\text{ bp}.
Fictional ten USD 1m parent orders: A receives five at 5 bp, B three at 4 bp and C two at 6 bp. Their contributions to aggregate cost are 2.5, 1.2 and 1.2 bp, totalling 4.9 bp or USD 4,900.
Fictional ten USD 1m parent orders: A receives five at 5 bp, B three at 4 bp and C two at 6 bp. Their contributions to aggregate cost are 2.5, 1.2 and 1.2 bp, totalling 4.9 bp or USD 4,900.
Original fictional allocation and cost ledger. Equal intended notionals and complete execution are stipulated; all costs use the same benchmark and fee convention. Target assignment probabilities happen to match realised shares in this sample.

The simple average across the three algorithms, (5+4+6)/3=5(5+4+6)/3=5 bp, gives each algorithm equal weight. It does not measure cost per dollar of this realised flow. With unequal parent notionals, even weighting by parent counts would be insufficient. For intended notional NjN_j and consistently defined parent cost cjc_j in bp, use

cˉ=∑jNjcj∑jNj.\bar c=\frac{\sum_j N_jc_j}{\sum_j N_j}.
Probability is not a quota. Under independent assignments with fixed probabilities, ten orders have expected counts of five, three and two. Those realised counts are not guaranteed. A weighted rotation designed to enforce quotas is a different assignment design. Moreover, order-count probabilities do not promise the same shares of notional when order sizes vary. Use the realised notionals for the cost ledger, rather than substituting the target wheel weights.

Why B's 4 bp is not a verdict

B has the lowest observed cost in the example. Yet it received only three parents, and nothing in the table establishes their difficulty or the uncertainty of the result. Perhaps they were liquid stocks traded during quiet periods, whereas A received urgent orders in wider spreads. That is confounding: the cost difference could reflect the flow rather than the broker algorithm.

Random assignment within a suitable eligible population can support a causal comparison by making assignment independent of potential execution outcomes. It balances difficulty in expectation, not exactly in every small sample. Predefined strata for liquidity, size and urgency can improve comparability; they do not remove sampling uncertainty. Restrictions that leave a broker eligible only for easy orders limit the comparison to conditions in which alternatives could actually have been assigned.

Keep assignment records and trader overrides. Dropping rejected or unfinished parents after routing introduces selection bias; treating every child fill as an independent experiment overstates the amount of evidence. A comparison must respect the assignment unit and dependence between orders exposed to the same stock or market session.

The benchmark can also change the ranking. Arrival-price slippage asks a different question from interval-VWAP slippage, and a favourable filled-price measure may hide poor completion. Complete-parent implementation shortfall can include the opportunity cost of unfilled quantity under a stated evaluation horizon. Choose the benchmark, fee treatment, completion measure and review window before seeing which algorithm looks cheapest.

The Robeco researchers' Optimal Order Routing with Reinforcement Learning illustrates weighted random routing for a particular segment of North American equity flow. Its evaluation uses VWAP-relative slippage scaled by spread, while discussing benchmark endogeneity and the need for additional metrics. That study's design should not be read as a claim that every wheel randomises all orders or that VWAP-relative cost is sufficient for every mandate.
Post-Trade Analysis develops benchmark reconciliation, notional aggregation and the limits of raw algorithm comparisons. The full lesson requires membership; the linked page provides a public preview.

Reweighting changes the experiment

A desk may change future allocation weights after reviewing transaction-cost analysis. The Robeco paper studies the resulting exploration–exploitation problem: allocating to apparently stronger brokers while retaining information about alternatives. It also notes that changing market conditions and broker improvements can distort pooled comparisons over time.

An operational review therefore needs dated policy versions, the eligible panel and probabilities at each assignment, pretrade characteristics, algorithm parameters, completion records and consistent benchmarks. A reweighting decision must use outcomes available at that time; tomorrow's completed trade cannot inform today's route. If weights adapt, the analysis must account for that assignment history rather than treating the entire sample as a fixed equal-probability trial.

A wheel makes allocation explicit and testable. Whether it improves execution still depends on its eligibility rules, evaluation design and subsequent evidence. A low average cost alone does not establish best execution.

Sources

Last updated
  • execution
  • broker allocation
  • transaction cost analysis
  • experimental design
  • algorithm selection
Keep connecting the ideas
← Back to the glossaryBack to top ↑