In personality psychology, Locus of control (Rotter 1954) is the degree to which people believe that they, as opposed to external forces, have control over the outcome of events in their lives. Individuals having high internal locus of control (henceforth LoC) perceive events as largely predictable and responsive to their actions (hence, controllable), and accordngly act in a more deliberate, confident, assertive and successful manner compared to individuals with a high external LoC. We seek to extend this individual level LoC analogy to the organizational level and examine whether and to what extent an organization's upper echelons perceive their decision making options as largely within their control versus dictated largely by external factors and pressures. We posit that such perceptions about locus of control in the firm's top management teams may well have systematic and measurable firm performance effects.

Whereas for individuals, the 29-item locus of control questionnaire scale (Rotter 1966) is widely used, the challenge with extending such a mental and behavioral concept to a firm's top management team in a firm performance context may well require a different measurement approach. We apply NLP (natural language programming) techniques on the language and vocabulary employed by top management teams in quarterly earnings calls to assess and extract an implied locus of control for the firm as a whole. We propose an "organizational locus of control" (henceforth OrgLoC) construct and measure it as the ratio of text units bearing an internal LoC orientation to those bearing an external LoC. We seek to investigate if OrgLoC impacts firm performance in measurable ways and if such impact is reflected in financial firm value.

The earnings call event is typically structured into a Prepared Remarks (PR) section delivered uninterrupted by the firm's top management team, followed by a question-&-answer (QNA) section with equity analysts. The two sections differ systematically in their structure and thereby, in the content and quality of information disclosed. To account for differential information content between the two sections of the earnings call, we categorize and analyze OrgLoC separately for each section (as OrgLoC_PR and OrgLoC_QNA).

Our text units of interest are 'relevant' sentences from the top management team, i.e. sentences pertaining to current and firm performance in any context (including decisions, actions, intent, expectations, contingencies etc.). To identify 'relevant' sentences, we first construct a manually labelled dataset of randomly drawn sentences from the earnings calls corpus as a training sample to train a battery of machine learning classifiers. We then assess the best performing classifier based on test sample accuracy and deploy the best performing model on the full corpus. The output of the exercise is a 'filtered' sub-corpus comprising 'relevant' public statements on current and expected future firm performance made by the organization's upper echelons to financial market participants in the course of quarterly earnings calls.

The challenge at this stage is classifying TMT statements as having intrinsic versus extrinsic (and sometimes neither) LoC orientation. This requires an analysis not just of the subject of each statement but also it's relation with the verb and object in each clause of each sentence. In that vein, we [1] use a grammar based tree parsing of each sentence, [2] identify SVO (subject-verb-object) tokens as our primary analysis objects at the token level, [3] detect and extract chunked phrases around each SVO token, [4] augment each SVO token with additional features (part-of-speech, morphological attributes etc.), and [5] employ these augmented SVO tokens in our analysis of LoC orientation of that sentence.

We first asked three independent coders to manually analyze a set of fifty sentences randomly drawn from the filtered corpus for the likely LoC orientation of each sentence into one of three classes - largely internal LoC orientation, largely external orientation and neither (cases in which coders either couldn't clearly pinpoint LoC orientation or couldn't agree on the same even after a discussion). We found high inter-coder concordance (kappa = x) and the distribution of sentences across the 3 classes (internal, external and neither) was 35%, 20% and 45% respectively. (*say) Then we next provided each coder a larger sample of 1000 sentences to be manually labeled into one of the above 3 classes. Then, we again follow the machine learning classifier model approached outlined previously, only that this time we are doing multi-class classification rather than binary class. We save the best performing model (logit model at 85% test sample accuracy) and apply it on the filtered full corpus. The output from this exercise is a distribution of LoC orientation for the PR and QNA sections of each earnings call for each firm. We define internal_ (external_)OrgLoC as the ratio of the number of internal_ (external_)LoC sentences to total relevant sentences, and OrgLoC simply as the ratio of internal_OrgLoC to external_OrgLoC.

As our outcome of interest for financial firm value, we choose the widely used cumulative abnormal returns (CAR) metric which summarizes the totality of a firm’s risk-adjusted prospects in the immediate to short term (and corresponds well with the time horizons for the methods we apply). We examine whether and to what extent OrgLoC impacts CAR after controlling for known determinants such as financial ratios, the Fama-French factors and earnings surprise (drawn from the Finance and accounting literatures), as well as standard controls such as firm and time fixed effects. We employ an event study approach followed by a calendar-time abnormal returns (CTAR) model with time-varying risk factors to investigate, validate, and empirically establish the (i) financial value-relevance, (ii) explanatory power and (iii) predictive power of organizational locus of control on cumulative abnormal stock returns over a quarterly horizon for a large sample of firms (S&P1500) over a 15-year time-frame. We find evidence that OrgLoC as a construct bears significant explanatory and predictive power over current and future abnormal returns.