decisions

October 14, 2023

process #

1. select buy candidates

2. for each candidate

   • indicators produce (multiple) predictions

3. prepare buy / sell (2nd level indicator)

   1. combine multiple predictions on each candidate (weight them)

   2. produce possible buy candidates

   3. evaluate current holdings

      for each holding, 2nd level indicator evaluate holding

      We have bought this holding with this prediction data Is the prediction still valid? can you (indicator) evaluate the current standing?

4. decide buy / sell

   With new candidates for buy, evaluations for current holdings

update process #

With prediction / evaluation, evaluate indicators / 2nd-level indicators

indicators #

decision unit

performs following (api) #

1. produce predictions

def predict(self, stock_id: str, current_price: Price) -> Prediction:
    # uses RawData
    pass

1. evaluate prediction on different time (how valid is the past prediction now?)

   if we made prediction on 3month granular, after 1 month, we’d be in position to evaluate that prediction

def evaluate(self, stock_id: price: Price, str, buy_data: BuyData, evaluate_at: datetime) -> Evaluation:
    pass

1. (supporting evaluation) how much I (indicator) can trust my previous prediction given data available after making the prediction

   I had so and so hypothesis when making predictions, and since they hypothesis turned out to be misplaced, I can re-evaluate my confidence on my previous prediction

2. (supporting evaluation) self check

   I just made this prediction, but I also know this depends on my assumptions In future, if this and this happens, I would advice differently.

2nd-level indicators #

aggregate multiple (selective) indicators

1. lv2 indicator produce 2nd level Prediction / Evaluation data

2. having access to multiple indicators can yield better decision

3. Whether to let indicator directly talk to another indicator is an implementation chice

data #

raw data for prediction #

raw data falls into following categories #

1. Target data

   Target data is what we are trying to get, but unable to get perfectly It answers how much an investment is worth and how much we are paying for

   • valuation (how much company value)
   • price (how much company cost (eg, stock price))

2. Target proxy data

   It is not directly related to the measurement of target data but it is a proxy

   • eg, company buying its own stocks can signal a new valuation

3. Environment data

   relavant data which affects interpretation of target data

   • inflation
   • interest rate

4. Speculation data

   • how people perceive market (people speculation)
   • how I perceive market (own speculation)

5. Past pattern data

   eg, MA trends analysis

Rawdata #

• data type
• data source
• data acquisition date

Prediction #

1. probability of going up / expected-amount (amount is in % unit)

2. when is this prediction about

   • prediction-horizon:
     • when is this prediction about
     • center and range (variance)

3. what is the nature of the prediction

   • valuation / speculation / environment

4. PredictionReason

   serves two purposes

   1. clearly represent the prediction logic when reviewing
   2. clearly indicate what the actual code ran
      • holds git commit hash and the function responsible for the prediction
      • might be better to place them where future execution run time can find so that we can readily use that code (create override function with commit hash)
      • when prediction / evaluation code is updated, it can indicate which previous versions it supercedes

5. FutureEvaluationReason

   similar to PredictionReason but for evaluation purpose

   1. clearly represent the evaluation logic at the time of prediction

In order to evaluate the prediction later, prediction has to show why (on what basis) we made the prediction by

• I predict Prediction based on RawData at access-time using PredictionReason
• and I think I can evaluate my prediction in future with FutureEvaluationReason

Evaluation #

similar to Prediction, but holds evaluation data

We have current holding at price different from prediction. We can evaluate current holding

1. by asking if we made wrong prediction in the past given we have more data available
   • it’s as if we are asking, this looks bad but did we really make wrong decision? or do we hold on to it?
   • It could be a simple rule to sell holdings
2. by producing new prediction with new data

Both 1, 2 can be beneficial

solely relying on 1 might miss a new opportunity solely relying on 2 might result in selling prematuarely

We also need to perform 1 for updating our PredictionReason / EvaluationReason

BuyData, SellData #

data stored when buying / selling takes place

BuyData #

1. purchase record (time, amount)
2. BuyDecision
   • 2nd level indicator’s buy decision data
     • list of [Prediction, weight]
     • 2nd level indicator’s DecisionReason

SellData #

1. selling record
2. SellDecision
   • list of [Evaluation, weight]

whether to buy #

expectation #

1. probability of going up
2. amount to win / lose

Indicator’s Prediction data is combined (by 2nd-level indicator) to create the expectation

how much to buy #

kelly #

F = P - (1-P) / R

F: % of investment (against total investment) P: probability to win R: expected-amount-to-win / expected-amount-to-lose

how to measure risk #

expectation - base (0) #

When expectation is high and above base, it’s low risk When investing long term, variance doesn’t matter much

variance #

higher fluctuation (of possible outcome) is higher risk initial portfolio theory used variance as risk measurement

consider #

high variance is mitigated if long term investment when equal expectation, lower variance is lower risk

how much to diversify #

when not prepared #

diversify, start from index fund

goal is to focus on relatively small # (5 or fewer) #

Current plan (ag1) #

decide if this is a good time to buy stock (index fund) so we have only single candidate

• decide which data to use

index fund indicator #

decide which raw data to use #

decide decision rule #