Behavioural Alpha from UK Retail Flows

The signal is derived from proprietary first-party behavioural data collected from TFE Group's financial calculator tools (ISA and Pension modelling applications). These are real user interactions, not survey data, panel estimates, or sentiment scraping.

Two event streams are ingested via the TFE Group API:

Each record is a user session capturing: provider selection (current and target), deposit amounts, wrapper type (Cash ISA vs Stocks & Shares), transfer intent, age demographics, and modelling parameters. Pension events capture: pot size, withdrawal age, drawdown strategy, annuity provider, and lump sum preferences.

Deduplication: Events are deduplicated by session_id to ensure one observation per unique user interaction. Users who refresh or adjust inputs are counted once.

Provider canonicalization: Provider names are normalised via a deterministic mapping dictionary (e.g., "HL", "Hargreaves", "Hargreaves Landsdown" all resolve to "Hargreaves Lansdown"). Unknown providers are matched using Levenshtein distance against a valid provider list (similarity threshold ~70%). Junk entries ("test", profanity) are filtered via regex.

Behavioural events are mapped to 11 LSE-listed equities via a static PROVIDER_TICKER_MAP. Multiple providers can map to the same ticker (e.g., Halifax, Lloyds, and Scottish Widows all map to LLOY.L).

Event classification per ticker per month:

Net signal formula:

The net signal is a share-of-flow measure, not an absolute count. A ticker with +5.2pp has 5.2 percentage points more buy-side flow share than sell-side flow share relative to total market activity that month. This normalisation controls for month-to-month variation in total sample size.

We evaluate three weighting approaches for the pension sell component:

All three methodologies are backtested against the same price data. The methodology with the highest directional accuracy is auto-selected as the production signal. Currently: ISA Only (w=0.0) at 71.4% accuracy.

The intuition: ISA transfer and modelling activity is a stronger leading indicator of provider sentiment than pension annuity selection, which is a one-time decision rather than an ongoing behavioural signal. Adding pension noise dilutes signal quality.

Time framing: Each month's behavioural data generates a signal vector. That signal is tested against the next available month's stock returns. In practice this is usually the following calendar month (e.g., January data predicts February direction). However, if there are gaps in the behavioural data (months with zero observations), the evaluation steps forward to the next month with data, which may extend the prediction horizon beyond one calendar month. This ensures the signal is genuinely predictive, not contemporaneous.

Directional accuracy:

Stock price source: Monthly closing prices from Yahoo Finance (query1.finance.yahoo.com, free, no API key, 6-hour cache). Returns are simple percentage change: (P[t] - P[t-1]) / P[t-1] * 100.

Missing price interpolation: If a month's close is unavailable, the return is estimated by finding the nearest available prices on either side and dividing the total return by the month span:

Incomplete month exclusion: If the outcome month is the current calendar month, it is excluded from all accuracy and correlation calculations because the full month's return is not yet finalised.

To prevent spurious signals from low-data months or thin-activity tickers, two minimum thresholds are enforced:

Signal magnitude filter: Only signals with |net_signal| >= 1pp are treated as directional predictions. Sub-1pp signals are classified as "no signal" (market-neutral for that ticker-month).

Confidence tiers based on signal magnitude:

Rank correlation (Spearman's ρ): Computed between net signal vectors and subsequent stock returns. We use Spearman rather than Pearson because signal magnitude is not linearly related to return magnitude. Only direction matters. Spearman tests monotonic rank-order agreement, which is the appropriate statistic for an ordinal directional signal.

Signal edge analysis ("Why Not Just Hold?"): All ticker-month observations are binned by signal presence:

This decomposition answers whether the signal adds value versus passive exposure. If buy-signal ticker-months outperform absent ticker-months, the behavioural data contains alpha that a "hold everything" strategy misses.

Signal returns are benchmarked against the iShares Core FTSE 100 ETF (ISF.L), the most liquid passive UK equity exposure.

This is a simple long-only comparison: if you bought only the stocks the signal flagged (equal weight) versus buying the FTSE 100 index, what is the excess return? The signal portfolio is rebalanced monthly based on new behavioural data.

The per-ticker signal engine generates directional predictions for 11 individual equities. The Aggregate Flow Index constructs an equal-weight portfolio from those predictions and tracks its performance against the FTSE 100 ETF (ISF.L).

This is a rebalanced monthly portfolio that goes long stocks where behavioural data shows net inflows (buy signals) and short stocks with net outflows (sell signals). When no sell signals exist (common with ISA Only methodology), the portfolio is long-only.

Alpha is measured as the excess return of this signal portfolio over the FTSE 100 ETF, answering: "If you only bought the stocks our signal flagged, would you outperform a passive UK equity allocation?"

A directional accuracy of 71.4% sounds compelling, but with only 49 testable predictions, how confident should you be that it is not random luck?

We run 10,000 Monte Carlo simulations where each simulation randomly guesses "up" or "down" for 49 predictions (fair coin, p=0.5). We then count how many simulations achieve an accuracy equal to or higher than the observed result.

Both the empirical (Monte Carlo) and analytical (z-test) p-values are reported. The z-score measures how many standard deviations the observed accuracy is above the random baseline of 50%. Conventional significance thresholds: p < 0.05 (moderate), p < 0.01 (high), p < 0.001 (very high).

The seed is fixed (seed=42) for reproducibility. Anyone with access to the count (35/49) can independently verify the p-value.

To independently verify or replicate the signal:

1. Obtain behavioural flow data via the institutional API access form below. We provide monthly provider-level buy/sell event counts for all 11 mapped tickers.
2. Fetch stock prices from Yahoo Finance (free). Monthly unadjusted close for each LSE ticker. We use raw close rather than adjusted close, so dividends are not reflected in returns. This is a known simplification.
3. Compute net signal per ticker per month using the formula above. Set pension_weight = 0 (ISA Only methodology).
4. Shift by one month: Compare month M signals against month M+1 returns. This is the core prediction lag.
5. Filter: Only include ticker-months with |net| >= 1pp and months with total events >= 20.
6. Score: Count directional hits: signal > 0 AND return > 0, or signal < 0 AND return < 0. Divide by total testable signals.

• Signal is directional only. The magnitude of net_signal does not predict the magnitude of stock returns. A +10pp signal is not necessarily better than a +3pp signal in terms of return size. Spearman rank correlation is used precisely because Pearson linear correlation is not appropriate here.
• 11-ticker universe. Coverage is limited to providers with publicly listed equity. Vanguard, Nutmeg, Moneybox, and other major platforms are not investable via this signal because they have no public stock ticker.
• Sample size. With 14 months and 11 tickers, the total testable observation count is 49. This is a small-sample backtest. Results should be interpreted with appropriate scepticism and are not a guarantee of future performance.
• Survivorship and selection bias. The ticker universe is defined by current provider-to-ticker mappings. Tickers that delisted during the sample period would be excluded, introducing potential survivorship bias.
• No transaction costs. Returns are gross of transaction costs, stamp duty, and bid-ask spread.
• Look-ahead risk. The signal strictly uses month M data to predict month M+1 returns. No future data leaks into the signal construction. However, the methodology selection (ISA Only vs weighted) is chosen ex-post based on full-sample accuracy. In live deployment, methodology would need to be fixed or selected via rolling window.
• Unadjusted close prices. Stock returns are calculated from raw closing prices, not dividend-adjusted close. This means dividend income is not captured, slightly understating total returns for high-yield tickers (e.g., LGEN.L, AV.L). Since the signal is directional, not magnitude-based, the impact on accuracy scoring is minimal.
• Yahoo Finance data gaps. When monthly close prices are missing, linear interpolation over the gap span is used. This smooths returns and may understate or overstate true monthly volatility.