AiLA Investment Strategies - A factor based approach

AiLA returns vs Big Three

Historical behavior
  • Overall low PnL correlation for AiLA vs 3x CTA, e.g., 𝜌(daily) ≲±15%.
  • AiLA captured bull market (“long trends”) from mid-2020 to Mar-2022, where Momentum strategies re-gained popularity (blue shaded period).
  • AiLA also captured crash (“short trends”) around Mar-2020, not captured well by slower CTA strategies.
  • AiLA typically also performed in other (“weak common trends”) periods.

AiLA’s Systematic Process

Historical behavior
  • AiLA alpha indices are opportunistic strategies using a systematic process.
  • The strategy is driven by both micro as well as macro factors.
  • Every step in the process is fully automated with no human intervention.

AiLA is Uncorrelated

AiLA vs CTA factors
  • 20 liquid commodity futures markets, front contracts.
  • 3 CTA strategies implemented based on [Baz15].
  • AiLA models aggregated for each commodity.
  • All strategies applied to same commodity returns.

[Baz15] J.Baz et al., Dissecting Investment Strategies in the Cross Section and Time Series (December 4, 2015). Available at SSRN:

AiLA Different to CTA Factors
  • Low return correlation between AiLA and the Momentum, Carry and Value strategies.
  • AiLA appears less related to any of the CTA strategies, than the CTA strategies among each other.
  • AiLA strategy returns are complementary to other systematic/quant strategies.
𝜌(daily) Momentum Carry Value AiLA
Momentum 1 62% -44% -2%
Carry 62% 1 -30% -13%
Value -44% -30% 1 -10%
AiLA -2% -13% -10% 1

AiLA Index Weights

  • AiLA products are defined by index weights, here referred to as the actual weights (AW).
  • The AiLA product weights are submitted via FTP at close of NY on T 1 for execution on market close of T.
  • The weight (AW) information consists of:
    • Asset (futures contract) name,
    • Weight in % of total allocated capital,
    • Position being long or short,
    • Rebalancing daily, to be executed at next close.
  • The weights are determined in downstream process, focusing on “how much” to allocate to each asset, given daily allocation decisions as input.

AiLA Weights Dissemination

Downstream Portfolio Construction

  • Inputs provided by the upstream process in the form of daily allocation decisions for individual assets.
  • On a given day, the inputs provide the decision to allocate long/short or not to allocate, for each asset in the AiLA index portfolio.
  • The downstream index construction process decide “how much” to allocate to each asset, by calculating index weights in three sequential steps based on the allocation inputs.

Downstream – Weights

Unconstrained Weights (UW)
  • Ideal weights are determined based on mean/variance portfolio construction principles, however, due to the dynamic nature of the allocations, only aspects empirically shown relevant for our circumstances are included in order to achieve a robust and transparent methodology.
  • The purpose of the unconstrained weights is to represent the ideal index, based on the user preferences before any practical constraints have been imposed. The following steps will then aim to achieve an index as close as possible to the unconstrained one while respecting the various user/execution constraints.
  • Unconstrained weights are calculated for the set of assets which have an active long/short allocation decision on a given date.
  • The unconstrained weights are calculated with the following aspects considered, of which some are optional.
    • Volatility: equally weighted w.r.t. asset risk.
    • N effective: varying allocation w.r.t. estimated effective number of assets.
    • Index correlation: induce a correlation preference to the standard index BCOM or GSCI.
    • Risk target: scale weights inline with index risk target.
Constrained Weights (CW)
  • The constrained weights are taking the unconstrained weights (UW) as an input and then determine the most similar set of weights given a set of linear constraints.
  • The calculation is only considering the assets with active (non-zero) weights on any given date.
  • The constrained weights are calculated with the following constraints.
    • Asset cap: upper bound w.r.t the individual asset weights.
    • Sector cap: upper bound w.r.t. the sum of weights within a sector.
    • Sharpe ratio: lower bound w.r.t. historical index Sharpe ratios.
    • Index allocation: equality bound w.r.t. the sum of all weights.
  • In case of a maximum allocation preference, an option is available to scale down the output weights to the max value on business days where the sum of weights exceeds this value.

Downstream - Allocations

Actual Weights (AW)
  • The index weights are calculated using the constrained weights (CW) as input.
  • In addition, a trade balancing period criteria is applied, where weights are set to zero when the number of business days to expiry or end-of-year is less than 2 (7) when using the Mid (Large) Cap execution logic.
  • The CW weights with the trade balancing period criteria applied is referred to as the target weights, which the actual index weights try to achieve.
  • However, the actual index weights are often not able to equal the target weights, due to the rebalancing restrictions implied by the different caps.
  • The different rebalancing restriction, respected by the execution logic when producing the actual weights, include user defined criteria as well as liquidity constraints dictated by volume/open-interest for the traded assets.
Downstream documentation

Further details about the downstream index methodology is found in,

Downstream Capital Allocation

Five asset index:

  • July Bean, June Copper, June Gold, Cal TTF and June WTI.
  • Portfolio target: equal 20% weight per asset, however, allow asset weights up to 100%.

Dynamic weights:

  • 2022 03 23: allocation opportunities for Gold, Copper and Soybean, weights increased above 20% resulting in system allocating 94.1% of capital.
  • 2022 05 16: TTF only asset with allocation opportunity, 20% weight prevented by liquidity constraints resulting in system allocating 6% of capital.
  • 2022 05 17: TTF allocation ends, resulting in system decision to exit 6% long TTF weight.

Varying number weights:

  • Periods with large (small) nr of active assets tend to result in smaller (larger) individual weights.

Upstream – Factors and Modeling

  • The main purpose of the upstream process is to decide “when” to allocate to the individual assets.
  • The allocations are intended to capture prevailing investment opportunities.
  • The opportunistic decisions are based on data driven models dedicated to individual assets.
  • The allocation decisions are produced by the upstream process in four sequential steps.

Factors affecting investment decision

  • Micro features are chosen specifically related to the asset and are entirely procured from data service providers Refinitiv and Bloomberg. The micro features data include but are not limited to curve structure data, commitment of traders as well as fundamental data.

    An example of micro features data would be H-J spread in RBOB which captures effect of temperature, RVP, expected refining capacity, expected post winter demand etc.
  • Macro features are used as common features for all assets. The data include but are not limited to market indices, FX exchange rates, economic indicators, and prices of key commodities. The data is procured entirely from data service providers Refinitiv and Bloomberg.
    An example of macro features data would be FX rate which has impact of product exports of RBOB from a refining rich location making it more or less expensive.
  • Feature data is collected after market close as part of the EOD process and all data is from repeatable, public and verifiable sources.

Upstream – Modeling Philosophy

Feature Data (XF)
  • All feature data is processed by the DQ system and then transformed into a format which allow the model to access relevant information more easily, e.g., including but not limited to data normalization, relative/spread information, one-hot-encodings etc.
  • The data is chronologically split up for the model training cycles, with three samples corresponding to consecutive periods (see chart),
    • Training: used to train model.
    • Validation: used to validate model performance as well as tune hyper-parameters.
    • Holdout: out-of-sample data to assess final performance after all model aspects have been frozen.
  • Due to the extraordinary risk of over-fitting, the holdout dataset can only be used ones for the final assessment of the model performance.
  • The most important features contributing to an allocation decision, associated with an asset on a given day, of any AiLA index are accessible in the AiLA product UI.

Upstream – Modeling Choice

Model choice: y = f(XF)
  • Model choice involves analysis of various models using the MLRD platform (Machine Learning Research and Development)
  • During the training-validation cycles of the models the hyper-parameters are also tuned.
  • The metrics collected, from the model run with specific emphasis on Accuracy and Signal Density form the basis of the qualitative judgement to determine the final choice of model.

Upstream – Explainable Factors

Allocation Decision (A)
  • Entry decision
    • The model output (y) each day provide a prediction w.r.t. achieving a risk-reward over a certain holding period.
    • For example, probability for a long position to achieve a 10% return, without loosing more than 3%, in next 15 days.
    • If probability sufficiently high for a favorable long or short risk-reward scenario, an opportunistic allocation will be entered with corresponding target parameters.
  • Exit decision:
    • An allocation will be kept until one of four exit criteria is met.
    • Expects little opportunity to be invested further.
    • Hitting the risk target since start of investment.
    • Achieves the reward target since start of investment.
    • Receiving an opposing opportunity from the model.
    • Above, only exit criteria (4) depends on the feature data.

Example: Gold_CME_June decision to allocate long.

2022-04-13: Entry decision

  • Favorable risk/reward predicted by micro/macro features
Close Price of Dec Gold 1 day prior to trade entry Date of Trade entry Close Price of Oct Gold 1 day prior to trade entry Close Price of May Gold 15 days prior to trade entry Open Price of Jan Gold 19 days prior to trade entry
Low Price of Oct Gold 2 days prior to trade entry Low Price of Apr Gold 9 days prior to trade entry High Price of Oct Gold 6 days prior to trade entry Low Price of Aug Gold 1 day prior to trade entry Low Price of Aug Gold 16 days prior to trade entry
US Weekly Domestic Crude Oil Production US Weekly Crude Oil Imports US Weekly Commercial Crude Oil Stock Dow Jones Index S&P 500 Index
10 Year Treasury Yield Index Bloomberg Commodity Index Baltic Dry Index USD Currency Index USD-JPY FX Spot
GBP-USD FX Spot CME Copper Front Month Gold Spot CFTC CoT NCNC Soybean Oil - CBT CFTC CoT NCNC #2 Heating Oil - NYMEX
CFTC CoT NCNC Natural Gas – NYMEX CFTC CoT NCNC Soybean Meal - CBT CFTC CoT NCNC Cotton No. 2 - ICE Futures U.S. CFTC CoT NCNC Live Cattle - CME CFTC CoT NCNC Feeder Cattle - CME
CFTC CoT NCNC Crude Oil - NYMEX CFTC CoT NCNC Cocoa - ICE Futures U.S. CFTC CoT NCNC Sugar No. 11 - ICE Futures U.S. CFTC CoT NCNC Coffee C - ICE Futures U.S. USD-BRL FX Spot

(* CFTC CoT NCNC = CFTC Commitments of Traders, Non-Commercial Net Contracts)

2022-04-21: Exit decision

  • Hitting risk target since start of investment