Research
Technical strategies, macro features and model-driven forecasts, with portfolio implementation examined alongside predictive evidence.
01 / Strategy research
TSM combines trend-following with relative value and shorter-term momentum across futures markets. Current extensions examine spread construction and an explicit FX overlay, with each component assessed on its own and within the combined portfolio.
Directional strategies
Directional signals across futures markets form the longest-running strand of the research. The focus is how trends at different horizons combine, how risk is allocated, and how the resulting exposures behave across market conditions.
Research extension
Relationships within asset classes provide a second source of trading ideas. Research examines spreads and butterflies, how hedge ratios are estimated, and when a relationship supports mean reversion or a continuing relative trend.
Technical strategies
Price breakouts and shorter-horizon continuation signals sit alongside the longer trend models. Studies examine their distinct contribution, signal persistence and the trade-off between faster response and higher turnover.
Overlay research
An explicit currency sleeve is being examined alongside the technical strategies, using six CME FX futures. The research considers its contribution to the combined portfolio, including overlapping positions, contract sizing and execution costs.
The components together
Trend-following, relative value, short-term momentum and FX are examined as one portfolio. The combined studies assess how the components interact, how they share risk and what remains after implementation costs.
Compare the standalone components with the combined portfolio, including shared exposures, changing correlations and behaviour through drawdowns.
Examine component sizing, risk concentration and the FX overlay within a common capital base. Portfolio outcomes depend on the allocation and capital assumptions used.
Net overlapping contracts, apply whole-lot sizing and trading costs, and assess turnover, margin demands and capital sensitivity.
02 / Signal & model research
A broad feature library brings together price behaviour, market structure and cross-asset relationships. Research tests how these inputs translate into forecasts across markets and horizons, and how model choices affect the information in those forecasts.
Trend, momentum, mean reversion, return distributions and intraday behaviour.
Curve level, slope and curvature; spreads, butterflies and relationships between markets.
Realised volatility, asymmetry, volatility ratios and cross-market risk relationships.
Volume, open interest, intraday participation and settlement-window measures.
Calendar effects, time to maturity, contract lifecycle and delivery-month structure.
Market-based indicators of growth, inflation, funding conditions, risk appetite and policy divergence.
Rolling statistical and machine-learning experiments span regularised regression, random forests and gradient-boosted trees. Studies compare training windows, feature groups and forecast horizons, with full-return and factor-controlled targets treated as distinct research problems.
Feature groups · Model experts · Evidence through time
Ridge expert studies examine predictive power by feature group, market and period. Each expert combines a feature group with a modelling configuration, training window and forecast horizon. Results are interpreted alongside the return target, preprocessing and regularisation used.
Compare feature-group forecasts across assets under matched modelling assumptions, examining predictive association, direction and payoff diagnostics.
Read cross-asset findingsExamine training windows, regularisation and target horizons, including dependence between experts and comparisons with simple return-mean controls.
Read expert-dependence findingsCompare pooled evidence with individual years and earlier history to assess persistence, changing relationships and sensitivity to the period studied.
Read longer-history findings03 / Efficacy & selection research
A forecast’s usefulness depends on when it is selected, how it is translated into exposure and how long that exposure is held. Studies examine efficacy across holding horizons, adaptive selection and forecast combination, using outcomes available at each decision point.
Frozen-holding studies examine the shape of forecast efficacy across maturities and how those relationships might inform holding decisions.
Studies compare selection based on recent shadow payoffs, IC and stability across feature-group experts and their model configurations.
Combination methods are compared with simple baselines to assess whether selection adds useful information after dependence and implementation are considered.
04 / Portfolio & capital research
The broader programme is working toward a coherent long/short implementation. Portfolio studies examine how forecast information translates into positions, how those positions share risk, and how implementation and capital requirements affect the result.
Exposures
Map outright and relative-value forecasts to underlying instruments. Separate residual-return diagnostics from actual portfolio returns, and net shared futures legs before assessing turnover and costs.
Construction
Compare forecast combinations, covariance-aware allocation, factor exposures and position constraints. Examine how holding rules and trading costs change the contribution of each signal.
Capital resilience
Study drawdowns, margin demands and leverage capacity across historical and block-bootstrap scenarios. Survival-capital estimates depend on the simulated conditions, risk model and implementation assumptions.
05 / Findings across the research
An index of selected studies supporting the research areas above. Each note identifies the question, study context, findings and remaining limitations, with links back to the relevant research topic.
Cross-asset studies · Residual targets
Training windows are compared across crude oil, equity, Treasury, currency and gold futures to examine responsiveness, persistence and the role of simple return-mean signals.
Finding. Shorter training windows can improve pooled predictive association without producing a consistent advantage within individual years or across markets. The useful horizon depends on the market and period.
Research implication. Compare fast and slow models on the same supported dates, retain simple controls and evaluate the resulting positions. A strong historical window is a candidate for further testing, not a selection rule by itself.
Cross-asset residual studies, September 2026. The reported residual payoffs are research diagnostics rather than executable futures returns.
Model behaviour · Dependence
Different feature groups and model settings can still produce closely related forecasts. Research examines what remains after accounting for common behaviour and aggregation.
Finding. Strong regularisation can pull forecasts toward a shared recent-return mean. Later equity comparisons find that lower Ridge penalties reduce overlap between group exposures, while substantial common movement remains.
Research implication. Model count is a poor guide to independent information. Forecast, signal-change and position comparisons help assess diversity; lower correlation alone does not establish a profitable combination.
CL expert-dependence study and cross-asset deep dives; subsequent equity expert comparisons, September–October 2026.
Equity studies · Full-return targets
A contrarian hypothesis is tested after weak results in the original forecast direction. The follow-up asks whether the apparent relationship survives a longer historical comparison.
Finding. Global forecast reversal produced a positive result in the inspected later sample. Extending the analysis into earlier history did not strongly replicate the portfolio result.
Research implication. The reversal remains exploratory. Fixed construction rules, normalisation history and differences in model vintages matter when interpreting replication. The positive later sample should not be read as a validated contrarian strategy.
Equity full-return and extended contrarian studies, October 2026. These are distinct from the earlier residual-target investigations.
Portfolio research · Implementation
A residual return, an outright return and a spread describe different exposures. Portfolio studies examine whether the predictive relationship can be captured in the instruments held.
Finding. Earlier residual studies show why predictive statistics and residual shadow payoffs need a separate implementation test. Later equity work maps spread signals into their underlying futures legs and applies costs after netting shared positions.
Research implication. Return definitions, factor exposures, execution timing, covariance and costs belong in the same experiment. Portfolio construction must be assessed alongside forecast quality; it cannot manufacture an edge absent from the inputs.
Cross-asset residual studies and subsequent full-return equity portfolio investigations, September–October 2026.