How we score every signal.

1. What a signal is

A signal is a computed observation about one security on one day, published as a non-personalised investment recommendation within the meaning of Article 20 of Regulation (EU) No 596/2014 on market abuse, with the disclosures it requires. It is not an opinion about the company and not a personalised recommendation: it is the result of applying a fixed procedure to closing prices, identical for every security in the universe and published to all subscribers at the same moment.

There are two families, and they do not mix:

• Bounce (mean reversion). A +3% target, a 10-session horizon and a maximum hold of 35 sessions. It looks for a security in a long-term uptrend (closing above its 200-session average) that has just fallen hard: its 2-session RSI drops below a threshold that adapts to the security’s own volatility (20, 15 or 10, from lower to higher ATR relative to price). It is not generated when the last two sessions traded more than 1.5 times their 20-session average volume, a sign of sustained selling.
• Impulse (trend following). A +10% target, a 180-session horizon and a maximum hold of 125 sessions: the horizon is the method’s reference duration, and the exit rules close any position before it is reached. It looks for a security already moving that may keep moving, and two generators produce it: momentum continuation (ADX above its own 80th percentile, with the close above its 21-session exponential average and that average above the 200-session one) and breakout confirmation (a close above the upper Bollinger band, on volume above its own 90th percentile and with the close above the 200-session average).

Each method publishes its own new signal, with its own date and levels. One company can carry a new signal of one method and an open position of the other at the same time, and an exit of one method does not affect the other method’s signal.

Every signal names its entry, its target, its stops and its exit rules in advance. The published entry is a limit buy order at a price equal to the close that produced the signal plus 0.5%, valid only for the next session. It fills at the open if the stock opens at or below that limit, and at the limit if the price touches it during the session; if it does not, the signal is discarded. The target and the stops are set on that published limit price, not on the price the order fills at, which is usually a little lower; the return is measured from the fill price. The entry is never the close the signal was calculated on.

Closing is mechanical too, and the full list is declared before opening: no exit is decided after the fact. During the session, a position closes when it reaches its target, on the volatility (ATR) stop or, for Impulse, on the fixed -20% stop. The ATR stop sits 4 ATR below the published entry, never further than -10%, for Bounce, and 6 ATR below, never further than -15%, for Impulse, and it acts only during the first 10 sessions for Bounce and the first 20 for Impulse; the Impulse fixed stop always acts. On the entry day, if the order filled at the limit during the session, the target only counts if the close reaches it. At the close the rules check overbought (2-session RSI above 80, for Bounce only and only with the position in profit), the tiered trailing stop (it activates from +0.5% for Bounce and +3% for Impulse, and acts only in profit), the loss of the 200-session average, stagnation (for Impulse, more than 60 sessions without ever gaining 2%), the time stop while the position is still at a loss (after 30 sessions for Bounce and 120 for Impulse) and the maximum hold (35 and 125 sessions), which closes the position in any case. Exits evaluated at the close are filled at the next open, so even an exit in profit can fill below the entry.

No new signal is published on a frozen price. When the ATR is below 0.35% of the price and the range of the last 10 sessions is below 1.5%, as happens when a cash takeover offer pins the share price, the published levels would mean nothing. The filter affects publication only: open positions and the backtest are unchanged.

Market context. Each market is classified every day into one of six states by its own index: its position against its 200- and 50-session averages and its one-month return. A new signal may carry a context label measured for its method, its market and that state (144 combinations): “very favourable” when the confidence interval of the mean return of past operations lies entirely above +0.40%, “favourable” when it lies entirely above zero, “no signal” when it crosses zero, “unfavourable” when it lies entirely below zero and “too few cases” with fewer than 30 operations. The interval groups operations by entry week, because those entering in the same week share the same market move. The label describes what happened in that context; it does not predict the signal, and it appears on new signals only. No market is vetoed: the state of the market withholds no signal.

2. Universe covered and inclusion criterion

Twelve stock indices are covered, around 1,300 unique securities; the exact figure, which changes with index membership and with the securities that have enough history, is the one shown at the top of this page. The inclusion criterion is objective and singular: membership of one of those indices. There is no discretionary selection, no watchlist of favourites and no exclusion by sector or by opinion.

United States: S&P 500, Nasdaq 100, Dow Jones and Russell 1000. Eurozone: DAX 40, EuroStoxx 50, CAC 40 and IBEX 35. Rest of Europe: FTSE 100, SMI, OMX 30 and OMXC 25.

A security needs at least 250 sessions of history before it can be scored; below that it stays out. Securities whose price series stops updating are withdrawn automatically before publication. Every week the membership of each universe is compared with its index and corporate actions (ticker changes, mergers, delistings) are reviewed.

3. Data sources and their reliability

Daily open, high, low, close and volume, plus fundamental data published by the companies themselves, obtained from external market-data providers, mainly Yahoo Finance. The download happens after each covered market closes, once its close is final at the provider, and the calculation runs on closing prices.

What does not enter the model or the signal rules, and it is worth saying:

• Intraday data, order book or market depth. The live session tracking subscribers see is informative only: the signals, their exits and their official record are calculated on daily bars.
• Analyst estimates, third-party research or consensus figures. The 360 analysis shows earnings per share against consensus as information; it enters no score.
• Alternative, proprietary or non-public data of any kind. Everything that enters the model is public market information.

Known limits: adjustment for dividends and splits depends on the provider and can arrive late; fundamental coverage runs between 92% and 99% depending on the index, and a security without it keeps its technical score but not its fundamental one. A series that has not updated in three days is treated as stale and that security is not published.

4. The model and its variables

Each family has its own model: a logistic regression estimating the probability that the signal closes in profit. The output is presented as a Conviction Score from 0 to 100.

Choosing a linear model is deliberate. It can be read in full, audited coefficient by coefficient, and its failure modes are visible. It was compared against non-linear gradient boosting alternatives and they did not beat it out of sample, so interpretability was not paid for with accuracy.

Twenty-three inputs: 21 technical and 2 regional. By family:

• Trend and position: distance to the 50- and 200-session moving averages, position and slope of the 21-session exponential average, MACD, ADX and the linear fit of the close over 10 sessions.
• Momentum and oscillators: 14- and 2-session RSI, stochastic, daily change and distance to the 20-session low.
• Volatility and range: ATR relative to price (ATR / close, which enters through two variables: relative ATR and relative volatility), and the position and width of the Bollinger bands.
• Volume: volume relative to its average, volume trend, and OBV against its average.
• Region: two indicators placing the security in the eurozone, the rest of Europe or the United States. They are there because the same technical indicators have behaved differently in each region.

Inputs are trimmed at their extremes, standardised and capped at three standard deviations, so one anomalous reading cannot drag the score. The Impulse model adds hinge terms on five variables, letting their effect change beyond a certain level; the Bounce model has none, because there they did not improve the out-of-sample measurement.

5. Training and validation

The models are trained on the simulated (backtest) operations of historical signals from the same universe, closed with the published exit rules: 114,368 Bounce signals and 53,472 Impulse. The deployed version is v10.24, trained on data through 2026-09-24.

Measurement uses a temporal split: the model is fitted on everything before a cut-off date and measured on what follows, with a purge zone at the boundary so that no training signal overlaps its horizon with the measurement window. The cut-offs are 2023-03-10 for Bounce and 2023-07-20 for Impulse.

Out-of-sample AUC is 0.599 for Bounce and 0.606 for Impulse. It is the measure of model quality we publish, never the in-sample one: the served model is then refitted on 100% of the history, so any measurement over that same data would be in-sample, would read higher, and would mean nothing. Measured on 24 September 2026, before the levels were set on the published entry (18 October 2026).

What an AUC is and is not. It is the probability that a signal which closed in profit scored higher than one which did not. 0.50 would be a coin toss and 1.00 would be always right. Our figures sit a good deal closer to the first than the second: they describe a modest, real ability to rank, not an ability to be correct, and ranking well is not beating the market (assumption 8).

6. Material assumptions

All of the above rests on assumptions that may not hold. These are the ones that carry the most weight, with what happens if they fail.

1. The past distribution of prices informs the near future

   This is the assumption holding up everything else. If the relationship between the indicators and the outcome breaks (a change of market regime, for instance) the scores keep being calculated and stop meaning what they say, with nothing failing visibly. That is why the model is monitored continuously, why a context label is measured by method, market and market state, and why changes are logged publicly.

2. The proposed entry is executable

   It is assumed that a limit order at the close plus 0.5% fills in the next session as described: at the open when the stock opens at or below the limit, at the limit when the price touches it. An opening gap, a thinly traded security or a wide spread can make the real entry differ from the published one, or not happen at all, and with it the whole result. The levels are set on the published limit price, so a fill below it does not move them.

3. Published figures are gross

   They exclude commissions, spread, taxes and currency conversion. This is not a detail: in our own simulation across account sizes from EUR 2,000 to EUR 250,000, a flat per-trade fee was enough to leave the Bounce strategy at zero or below at every account size. The deciding variable is the broker, not the size of the account.

4. No future information enters the calculation

   History is reconstructed under a time lock: each signal is scored only with what existed that day. This assumption already failed us once, with metrics built on a snapshot that contained later data, and that is why track-record badges are presented as historical context and never as predictive power.

5. The historical universe carries survivorship bias

   History is built from the current composition of the indices, so securities that left them are under-represented. The effect pushes historical figures upward, and its order of magnitude has been estimated: in the US, an equal-weight portfolio of the current members returns 15.1% a year over twelve years, against 12.6% for the index (SPY), before tax, and part of that gap is the bias. Every historical figure is therefore a ceiling.

6. The model is stable between retrainings

   There is no online learning: coefficients are fixed at retraining and do not change until the next one. The further the market moves from the training period, the weaker the guarantee behind the published measurement.

7. The score measures probability, not magnitude

   The model estimates whether a signal closes in profit, not how much is gained, nor how much is lost when it does not. A high score does not describe the size of the possible loss.

8. Closing in profit is not beating the market

   Conviction estimates whether a signal closes in profit, not whether it does better than its own index over the same days. Measured that way, out of sample (25 September 2026, before the levels were set on the published entry), the signals’ edge over the index is zero for Bounce (-0.01% per operation) and negative for Impulse (-0.19%), gross and before tax. No figure on this page claims or suggests that the methods beat the market.

7. Risk and sensitivity

Investing in equities involves the risk of loss, including the total loss of the capital invested. The signals, rankings and scores are non-personalised investment recommendations within the meaning of Article 20 of Regulation (EU) No 596/2014 on market abuse, published with the disclosures it requires; the rest of the content of this platform is general market information. None of it is investment advice or a personalised recommendation. Past results do not guarantee future results.

On the sensitivity of the score, which usually goes unsaid. Measured on 24 September 2026, before the levels were set on the published entry (18 October 2026): out of sample, for Bounce the low conviction band closed in profit 60.8% of the time for a +0.26% mean, the medium band 72.6% (+0.28%) and the high band 79.0% (+0.49%); for Impulse, 54.0% (+0.35%), 70.7% (+1.16%) and 73.7% (+1.17%). These are gross backtest figures, before tax. The conviction bands are different products wearing one face, and the difference between adjacent bands is a few points.

As a result, a small change in the input data (a revised price, a dividend arriving late) can move a score from one band into the next. The boundary between bands is not a line between good and bad: it is a cut across a continuous variable.